ArticlePDF Available

A Game of Droid and Mouse: The Threat of Split-Personality Malware on Android

Authors:

Abstract

In the work at hand, we first demonstrate that Android malware can bypass current automated analysis systems, including AV solutions, mobile sandboxes, and the Google Bouncer. A tool called Sand-Finger allowed us to fingerprint Android-based analysis systems. By analyzing the fingerprints of ten unique analysis environments from different vendors, we were able to find characteristics in which all tested environments differ from actual hardware. Depending on the availability of an analysis system, malware can either behave benignly or load malicious code dynamically at runtime. We also have investigated the widespread of dynamic code loading among benign and malicious apps, and found that malicious apps make use of this technique more often. About one third out of 14,885 malware samples we analyzed was found to dynamically load and execute code. To hide malicious code from analysis, it can be loaded from encrypted assets or via network connections. As we show, however, even dynamic scripts which call existing functions enable an attacker to execute arbitrary code. To demonstrate the effectiveness of both dynamic code and script loading, we create proof-of-concept malware that surpasses up-to-date malware scanners for Android and show that known samples can enter the Google Play Store by modifying them only slightly.
A Game of Droid and Mouse:
The Threat of Split-Personality Malware on Android
Dominik Maier, Mykola Protsenko, Tilo M¨uller
Department of Computer Science
Friedrich-Alexander University Erlangen-N¨urnberg
Germany
Abstract
In the work at hand, we first demonstrate that Android malware can bypass
current automated analysis systems, including AV solutions, mobile sandboxes,
and the Google Bouncer. A tool called Sand-Finger allowed us to fingerprint
Android-based analysis systems. By analyzing the fingerprints of ten unique
analysis environments from different vendors, we were able to find characteristics
in which all tested environments differ from actual hardware. Depending on
the availability of an analysis system, malware can either behave benignly or
load malicious code dynamically at runtime. We also have investigated the
widespread of dynamic code loading among benign and malicious apps, and
found that malicious apps make use of this technique more often. About one
third out of 14,885 malware samples we analyzed was found to dynamically load
and execute code. To hide malicious code from analysis, it can be loaded from
encrypted assets or via network connections. As we show, however, even dynamic
scripts which call existing functions enable an attacker to execute arbitrary code.
To demonstrate the effectiveness of both dynamic code and script loading, we
create proof-of-concept malware that surpasses up-to-date malware scanners
for Android and show that known samples can enter the Google Play Store by
modifying them only slightly.
Keywords: Android Malware, Dynamic Code Loading, Dynamic Script Loading
Email addresses: dominik.maier@fau.de (Dominik Maier), mykola.protsenko@fau.de
(Mykola Protsenko), tilo.mueller@fau.de (Tilo M¨uller)
Preprint submitted to Computers and Security April 20, 2015
1. Introduction
The processing power of current smartphones already surpasses that of last
decades personal computers. When they became a lot faster, our usage pattern
quickly shifted away from solely placing phone calls. Today, we do not only
store our family, friends, and business contacts on smartphones, but also check
for emails, visit social networks, and access bank accounts to transfer money.
Both, our private and business life, takes place on smartphones. By this, we
entrust our personal data to them, including photos, movement profiles, and
text messages.
As of today, the largest operating system for smartphones is Android with
a market share of 81.9% [
1
]. This makes Android an ideal target for shady
IT practices. Android versions on older phones do not get updated quickly,
sometimes not at all. As a consequence, publicly available exploits can be
targeted towards a large number of users for a long period of time. The Android
Market, which is today re-branded to Google Play Store, has had a lot of malware
uploaded in the past. Besides the Play Store, the open source nature of Android
allows vendors to ship their devices with markets from other companies, like
the Amazon Appstore, and also holds an option for users to install apps directly
from the Internet.
1.1. Contributions
We investigated Android malware that behaves benignly in the presence of
analysis, and maliciously otherwise. Automated static analysis scans all code
that is locally present in cleartext, while dynamic analysis executes the code
and observes its behavior. To surpass both analysis techniques, we exploit
split-personality [
2
] attacks that first split malware samples into benign and
malicious parts, and then mislead security scanners by showing only benign
behavior within analysis environments.
In detail, our contributions are as follows:
2
We study the use of dynamic code execution techniques in malicious and
benign apps by analyzing a set of 14,885 malicious and 22,032 benign apps.
From the malicious set 36.4% of the samples use code loading, while from
the benign set only 13.1% use code loading.
As part of our work, also a fingerprinter for Android sandboxes has been
developed called Sand-Finger. With Sand-Finger, we were able to gather
information about ten Android sandboxes and AV scanners which are
available today, such as BitDefender and the Mobile Sandbox [3].
With the results of Sand-Finger, we prove that Android malware can easily
gather information to surpass current AV engines by split-personality
behavior. We provide proof-of-concept samples that use a combination of
fingerprinting and dynamic code loading to enter the Google Play Store.
We show that an attacker can, even in an environment that forbids to
dynamically load compiled code, still run arbitrary statements by using
scripting environments like Android’s v8 JavaScript-VM to execute code
that interacts with the Android OS. In a proof-of-concept, we show it is
possible to run most statements in the Android API by combining dynamic
script loading attacks with Java’s reflection API.
Finally, we give an overview about sandbox improvements and counter-
measures that have been previously proposed to help prevent such attacks.
We argue, however, why these improvements are difficult to implement on
Android and why malware detection generally fails in practice.
To sum up, we show that split personality malware based on dynamic code
loading is a real threat that can effectively surpasses all security mechanisms for
Android, including sandboxes, the Google Bouncer and on-device AV scanners.
1.2. Related Work
How to surpass virtualization-based sandboxes on x86 has been shown by
Chen et al. in 2008 [
4
]. A comprehensive security assessment of the Android
3
platform has been given by Shabtei et al. in 2010 [
5
]. The security of automated
static analysis on Android, especially concerning virus scanners, has recently
been evaluated in 2013, for example, by Rastogi et al. [
6
], Fedler et al. [
7
],
and Protsenko and M¨uller [
8
]. However, none of these approaches focus on
bypassing dynamic analysis of automated Android sandboxes. The bytecode-
based transformation system PANDORA [
8
], for example, complicates static
analysis by applying random obfuscations to malware, but cannot surpass
dynamic analysis as it does not change its behavior.
From an attackers point of view, semantics-preserving obfuscation that con-
fuses static analysis is outdated due to the increasing use of Android sandboxes,
like the Mobile Sandbox [
3
], and at the latest since the initial announcement
of the Google Bouncer in 2012 [
9
]. A short time after the announcement of
the Bouncer, it was proven that dynamic analysis cannot be entirely secure by
Percoco and Schulte [
10
], as well as Oberheide and Miller [
11
]. Most recently,
in 2014, Poeplau et al. have shown that dynamic code loading can bypass
the Google Bouncer [
12
]. Poeplau et al. come to the conclusion that dynamic
code loading is unsafe in general and must be secured on Android. However,
obfuscation and packing techniques that bypass all Android sandboxes and AV
scanners have not been investigated in a large-scale study yet.
The paper by Poeplau et al. specifically centers around dynamic code loading.
They come to the conclusion that dynamic code loading is unsafe and should be
secured on Android. They also propose a possible solution that we evaluate in
this paper [
12
]. The other evaluated solutions to secure dynamic code loading,
targeted mainly towards native code, have been released by Fedler et al. [
13
,
14
,
7
].
A comprehensive security assessment of the Android platform has been given
by Shabtai et al. [
5
]. Short time after the initial announcement of Android
Bouncer, a new security layer on Android it was proven that it is not 100%
secure. This conclusion was mainly underpinned by the work “Adventures in
Bouncerland” from Percoco and Schulte [
10
] as well as the talk at SummerCon
2012 by Oberheide and Miller [
11
]. They fingerprinted the Android Bouncer
and used dynamic code loading to successfully submit malicious payloads to the
4
Android Play Store.
Fingerprinters that track Android malware sandboxes to be used in split-
personality malware have been proposed and their use against those sandboxes
has been proven by different researchers [
15
,
16
,
17
]. Luo et al. [
18
] have suggested
attackers might exploit a WebView inside a previously prepared app. In addition
to the Android world, research on how to evade VMs has been conducted on
Windows for years, such as [
19
]. Loading malicious code dynamically as a
scripting language has also been used by malware in the wild. Research on the
“Flame” sample has been conducted and released by Bencs´ath et al [20].
1.3. Outline
The remainder of this paper is structured as follows: In Sect. 2, we provide
necessary background information for available anti-malware technologies on
Android, such as AV scanners and sandboxes. In Sect. 3, we overview the
widespread of dynamic code execution techniques in malicious and benign apps
in the wild. In Sect. 4, we describe our own approach to fingerprint analysis
environments with the tool Sand-Finger. In Sect. 5, we show that Android
malware can easily bypass automated security scanners by acting according to
the principle of code splitting. In Sect. 6, we show how countermeasures that
prohibit dynamic code loading can be tackled by dynamic script loading in the
future. In Sect. 7, we discuss how this problem can again be tackled by looking
at more sophisticated countermeasures. Finally, in Sect. 8, we conclude that
dynamic code loading attacks yield an arms race on Android, or a game of
cat-and-mouse.
2. Background Information
We now provide background information about automated malware analysis
for Android. In Sect. 2.1, we illustrate static analysis, pointing out some
advantages and problems. In Sect. 2.2, we focus on dynamic analysis.
5
2.1. Static Analysis
Static analysis is the art of analyzing code without actually running it. Static
analysis tries to find out the control flow of an application in order to figure
out its behavior [
21
]. It is also often applied for finding programming faults
and vulnerabilities in benign applications, and is furthermore used to detect
malicious behavior without having to run an application. This paper uses the
term static analysis for every analysis that is performed statically as opposed to
running code dynamically. Note that static analysis is often understood as the
task of manually analyzing malware, which is very time consuming and requires
specially trained analysts. The principles discussed in this paper cannot defeat
manual analysis entirely. However, as it is nearly impossible to perform manual
analysis for all apps that are uploaded to the Play Store, we confine ourselves to
automated analysis.
Currently, virus scanners on Android are not allowed to monitor running
apps, but can only statically analyze them. Due to these restrictions, monitoring
an app’s behavior on real devices is close to impossible. Tests show that static
virus scanners on Android can easily be evaded; proof has been released by
various authors [
8
,
6
,
7
,
12
]. For manual static analysis, sophisticated tools
exist for both DEX bytecode and compiled ARM binaries. Contrary to that,
considering automated static analysis on Android, native binaries are often black
boxes for malware scanners since the tools focus on DEX bytecode. In native
code, however, root exploits can open the doors for dangerous attacks.
2.2. Dynamic Analysis
As a solution to the shortcomings in automated static analysis, running code
in a sandboxed environment received attention. By running code instead of
analyzing it, automated tools can trace what an app really does. The logging
files can then be rated automatically, or be examined manually. Although
approaches for running software on bare-metal exist [
22
], dynamic analysis of
mobile apps is almost exclusively done on virtualized environments. Virtual
machines considerably simplify the design of sandboxes because (1) changes in a
6
VM state can be monitored by the host system, and (2) a VM can be reset to
its initial state after an analysis.
Sandboxes for desktop malware have existed for years, including the CWSand-
box proposed by Willems et al. [
23
] as well as Anubis [
24
]. With the rise of smart-
phones, sandboxes for mobile systems were also proposed at that time [
25
,
26
].
After the introduction of touch-centric OSs for smartphones, in particular iOS
and Android, old smartphone systems quickly declined in sales and researchers
started to adopt sandboxes for the new platforms. On the one hand, the openness
of Android leaves the possibility to spread malware more easily, but on the other
hand, thanks to this openness, it has proven to be easier to implement a sandbox
for Android than for iOS [
27
]. Recent sandbox technologies are mixing static and
dynamic analysis for the best results. As shown in our comparison of sandboxes
in Sect. 4.2, many sandboxes also use AV vendors to detect Android malware.
3. Dynamic Code Loading in the Wild
The widespread of dynamic code loading techniques was first studied by
Poeplau et al. [
12
] for benign apps. They outlined five mechanisms of loading
and executing additional code, previously not included in the original APK.
These mechanisms are class loaders, package contexts, the use of
Runtime
class,
increase of the code basis through native code, and installation of additional
APKs. They have statically analyzed over 1,700 benign apps and found one
third of them to be using one of these code loading techniques. Here, we have
conducted a similar study that is limited to the most important mechanisms
class loaders and the use of
Runtime
class, but compares a larger dataset of
14,885 malware samples and 22,032 benign apps.
For our evaluation, we implemented an analysis tool based on the Soot
framework [
28
] which is able to directly input from
.apk
files and perform
various custom analyses on the intermediate representation without requiring
the source code of an app. Dynamic code execution techniques considered in
this study, as well as our approach to their detection are briefly explained below:
7
Class Loaders allow the app to dynamically load additional classes from
various file formats and locations. We recognize an app to be using this
technique if it calls a method with the name
’loadClass’
of some class
with the name ending with ’ClassLoader’, e.g., DexClassLoader.
exec()-Method of the Runtime class allows execution of an arbitrary bi-
nary file. Detecting this feature can be done by searching the bytecode
instructions of an app for the invocation of the
exec()
method from the
Runtime class.
The results of this investigation are presented in Figure 1. Summarizing, 36.4%
of the malicous samples and 13.1% of the benign apps were detected to be using
dynamic code loading. Note that these ratios are lower bounds because we
searched for the most frequently used code loading mechanisms but may have
missed apps that use another technique. Also note that direct processing of
Android bytecode in Soot is a new feature under active development which is
not always stable. Hence, the number of analysis failures is shown in the third
row of the table.
Malicious Benign
Samples total 14885 22032
Analysis Failed 2350 117
Processed Successfully 12535 21915
Class loading 1118 (8.9%) 1546 (7.1%)
Runtime class 3769 (30.1%) 1675 (7.6%)
Both 329 (2.6%) 358 (1.6%)
Either 4558 (36.4%) 2863 (13.1%)
Figure 1: Widespread of dynamic code loading in Android applications.
The results are consistent with those obtained by Poeplau et al. and show
that dynamic code loading is not a rare occurrence, which indicates the relevance
of the problems brought to discussion in this paper. Furthermore, one can
8
see that malicious apps make use of dynamic code execution generally more
frequently than benign ones.
4. Fingerprinting Analysis Environments
In this section, we introduce our fingerprinting tool Sand-Finger (Sect. 4.1),
and give a comparative overview of available sandboxes (Sect. 4.2).
4.1. The Sandbox Fingerprinter
Smartphones are usually embedded devices with built-in hardware that
cannot be replaced and, as opposed to desktop PCs, they are aware of their
specific hardware environment. Many different kernel and Android versions
exist with various features for different hardware. Additionally, smartphones
are actively being used and are not run as headless servers. In other words,
users install and start different apps, interact with the GUI, lock and unlock
the screen, move the device, and trigger numerous other system events. All
this device information, including the specific hardware, location data, and user
generated input, can be exploited to differentiate systems, which are used by
humans, from automated analysis environments. Hence, this information can be
exploited to detect dynamic analysis environments.
During the development of the Sand-Finger fingerprinter, we collected as
much data as possible from sandboxed environments and submitted it to a
server for evaluation. Building a VM close to actual hardware is a sophisticated
task due to the large number of sensors, e.g., accelerometer, gyroscope, camera,
microphone, and GPS. In short, we could not find any sandbox with hardware
that is entirely indistinguishable from real hardware. In addition to hardware,
malware can read settings and connections to find out whether it is running on
a real device or not. For example, constraints like “the device needs at least
n
saved WiFi-networks”, “the device must have a paired Bluetooth device”,
“the device must not be connected via cable”, “the device IP must not match a
specific IP range”, and “the device must have at least
n
MB of free space”, can
9
easily be put in place and require dynamic analysis tools to emulate realistic
environments.
Note that the examples given above are not far-fetched, because at the
moment no emulator we analyzed has an active or saved WiFi connection. Also
note that detection rates with a small number of false positives do not prevent
malware from being spread. To get accurate fingerprint results, additional
oddities can be taken into account. For example, the default Android emulator
does not forward the ICMP protocol used by ping. Whenever an obvious problem,
like the empty WiFi list or missing ICMP forwarding, gets fixed, malware authors
will come up with new ideas. Instead, static analysis must be used to reach
high code coverage and to skip the fingerprinting part of malicious apps [
29
].
However, according to our results, none of the tested analysis environments can
fool simple fingerprinting and reach high control flow coverage.
The app that we developed for this work, namely Sand-Finger, sends an
automatically generated JSON-encoded fingerprint to a remote web service via
HTTPS. JSON (JavaScript Object Notation) is a human-readable data format for
which many parsers exist. HTTPS is used to prevent the filtration of fingerprint
data. While performing our tests, we noticed that some requests were filtered
and never reached our server. Some of these sandboxes, however, could then be
assessed without encryption.
When executed in a sandbox, the fingerprinter gathers information available
on the virtual device. Around 100 unique values are extracted, depending on
the platform version. The information includes static properties like the device
“brand”, “model”, “hardware”, “host name”, “http agent”, and a value named
“fingerprint”. Information about the underlying Linux system is extracted as
well, such as “Kernel version”, the “user”, and the “CPU/ABI”. On top of this
static information, a lot of device settings are gathered, such as the timezone,
the list of WiFi connections, and whether debugging is enabled. Additional
data is gathered from the surroundings, namely the current phone’s cell network
status and neighboring cell info. Our analysis also checks if the device has been
rooted by searching for the su binary. Moreover, it lists the directories of the
10
SD card and the system’s bin-folder and transmits all available timestamps.
According to our tests, it is very useful to know how long a system has
already been running, since analysis systems are often reset. We also send a
watermark back to our server, which is a constant string that is specified at the
compile time of Sand-Finger to identify the origin of each test run. Thanks to
this, the redistribution of malware samples between sandboxes and AV vendors
can be traced. For example, the Mobile Sandbox passes on suspicious samples
to companies like BitDefender and Trend Micro, if a sample was not marked as
private.
Last but not least, two pings to
google.com
and our own server are performed
on the command line. Since the QEMU VM, in its default mode for the Android
emulator, does not forward packages other than UDP and TCP, an app is
presumably inside the default emulator if no ping reply is received. Hence,
the values returned by ping are also added to our JSON-fingerprint. After a
complete fingerprint is received by the server, it is stored inside a database
together with the server’s timestamp and the sender’s IP address. Sand-Finger
can also embody malware, that combines dynamic code loading against static
analysis with a split-personality behavior, making it invisible to automated static
and dynamic testing. The client listens for server responses, and only if an
expected response is received, does Sand-Finger extract its payload and run it
using System.load().
4.2. Comparing Android Sandboxes
In this section, we give an overview about available Android sandboxes and
AV scanners, and compare these analysis environments based on the results
of our fingerprinter Sand-Finger. Anti-virus companies often receive samples
from sandboxes; for example, according to the identification of our watermarks,
the Mobile Sandbox forwards samples to several online AV scanners, including
BitDefender and Trend Micro. The IP of one fingerprint we received for the
watermarked Sand-Finger sample that was submitted to the Mobile Sandbox
belongs to the anti-virus company BitDefender, while another was sent from a
11
system with the model TrendMicroMarsAndroidSandbox. Contrary to that, the
Google Bouncer does not forward submitted samples to AV scanners. Moreover,
the Google Bouncer does not seem to connect to any internet service, including
our web server waiting for fingerprints. Consequently, we could not reliably
fingerprint the Google Bouncer (but we were still able to surpass it with simple
split-personality malware, as explained in Sect. 5.2).
As the JSON file generated by Sand-Finger has over 100 entries to identify a
system, we focused on values in which all analysis environments greatly differ
from real hardware. Altogether, we found about ten of such values, as listed in
Fig. 2. The hardware we used as a reference value is the Nexus 5 from Google,
but we confirmed our results on other hardware, too. As a second reference
value, we used the official emulator of the Android SDK. The meaning of the
columns in Fig. 2 is:
Android Version: The version number of Android.
CPU Architecture: The actual CPU architecture. In some cases, the
CPU/ABI differs from the os.arch type in Java, so the latter was used.
Resembles Hardware: Some sandboxes try to adapt the virtualization
environment to resemble real devices, mostly Google Nexus. However,
others just use stock values or something completely different.
Successful Ping : The official Android emulator cannot forward pings, only
UDP and TCP packets are routed.
Correct Timestamp : The timestamp in virtualized environments is of-
tentimes inaccurate, but accurate timestamps can be received from the
Internet. This column lists if the timestamp is close to correct.
Hardware not Goldfish : The virtual hardware of the Android emulator is
called Goldfish. This column lists if the tested environment uses a different
hardware.
12
WiFi Connected : A real device is usually connected to the Internet either
via cell-tower or WiFi, otherwise the fingerprint could not reach our server.
Sandboxes are often connected via cables instead.
Cell Info: If the device is a real phone, it has some kind of cell info
describing the current mobile connection. Sandboxes often omit this info.
UpTime
>
10 min.: As a normal phone or tablet runs the whole day, the
probability of a device running longer than 10 minutes is high. Sandboxes
are often reset after each analysis.
Consistent: A consistent fingerprint from a real device does not show two
different Android versions, or two different CPU architectures.
Untraceable: A sandbox that gives away its own location or its IP address
in its fingerprint is traceable; otherwise it is untraceable.
In the following, we discuss the results of ten fingerprinted sandboxes and AV
scanners in alphabetical order: Andrubis,BitDefender,ForeSafe,Joe Sandbox
Mobile,Mobile Sandbox,SandDroid,TraceDroid,Trend Micro, and two systems
from unknown vendors. The unknown systems stem from the redistribution of
malware samples among sandboxes. We received several fingerprints from these
sandboxes, but could not identify their operator.
Andrubis
This is the mobile version of Anubis, a well-known sandbox for desktop
PCs [
24
]. Andrubis performs static as well as dynamic analyses and allows
uploading APK files. It also offers an Android app to submit APK files. A
problem about the APK files, however, is the size restriction of seven MB [
30
].
Andrubis has an extensive feature set, including support of the native code
loading. Like many other sandboxes, it is based on open source projects like
DroidBox,TaintDroid, and Androguard [
31
]. The Andrubis sandbox, despite its
usefulness when it comes to detection of malware, can easily be fingerprinted.
13
Its environment is similar to the official emulator, and it even uses test keys
(meaning it is self-signed). Another drawback is, that it still runs on Android
version 2.3.4, probably because of its dependency on the popular analysis software
TaintDroid [
32
]. As on most sandboxes, also on Andrubis, a ping did not reach
its destination. The static IP address used by Andrubis ends at the Faculty
of Informatics of the TU Vienna. Regarding this sandbox, we noticed strange
behavior when an app closes: Andrubis just taps on the desktop wildly, launching
the music app or executing a Google login, for example.
BitDefender
In the few weeks of our analysis, while we were working on the fingerprinter,
the BitDefender sandbox apparently was updated. First, the VM was inconsistent,
actually being a 4.1 Android that claims to be 4.4. Then the system changed,
and since then consistently shows Android 4.3. This makes it the most recent
Android version among the group of tested sandboxes. However, BitDefender is
not the stealthiest environment because values like “Batmobile” for manufacturer
and “Batwing” for the product can easily be fingerprinted. But note that as long
as nobody fingerprints these values, they are more secure than the stock names
from the Android emulator, because it is impossible to white-list all Android
devices. Interestingly, BitDefender was the only sandbox that contained cell
information to emulate a real phone. The static IP address used by this sandbox
is registered for BitDefender.
ForeSafe
The ForeSafe sandbox not only offers a web upload but also a mobile security
app for Android. After analysis, ForSafe shows an animation of the different
UI views it passed while testing. It is not the only Sandbox that captures
screenshots, but the only one with actually moving screenshots [
33
]. ForeSafe
uses Android 2.3.4 and hence, probably builds on TaintDroid. Despite heavy
fingerprinting and the possibility to execute native code, Sand-Finger was flagged
as benign by this scanner. ForeSafe executed the sample a few times and was
14
the only scanner with an uptime exceeding ten minutes. This is important as
malware could simply wait longer than ten minutes before showing malicious
behavior.
Joe Sandbox Mobile
This is a commercial sandbox which has a free version as well. Just like
the sandboxes above, it is possible to upload APK files for automated testing.
According to our tests, this sandbox performs very extensive tests, combining
static and dynamic analysis, and it even detects some anti-VM techniques in
malware. Roughly speaking, it is the most exotic sandbox in our test set. First,
we found out that Joe Sandbox uses the anonymous Tor network as proxy and is
thereby untraceable. Second, it is the only sandbox that does not build upon the
official Android emulator, but uses an x86 build of Android for the Asus Eee PC.
In addition to the x86/ABI, Joe Sandbox specifies “armeabi” as “CPU2/ABI”,
meaning that Intel’s libhoudini.so is probably installed to run binary translated
ARM code. Most likely, Joe Sandbox Mobile is running on VMware; the build
settings, however, are adapted to mimic a Galaxy Nexus. As it is based on a
different virtualization technology, pings can reach their destination. Moreover,
the build is rooted. The sandbox itself was generally hard to trick, e.g., some
anti-virtualization techniques are detected.
Mobile Sandbox
Another advanced sandbox with many features in the making, like machine
learning algorithms, is the Mobile Sandbox. It offers a web platform to submit
malware samples and is based on a modified version of DroidBox for dynamic
analysis. It also applies static analysis and monitors usage of native libraries as
well as network traffic. The camouflage of Mobile Sandbox is the most advanced
within our test set. Just by looking at its properties, it is indistinguishable from
a Samsung Nexus S, even replacing the Goldfish hardware string with a real
device string. Moreover, a telephone number was given, and it has a more recent
Android version than most sandboxes we tested. Interestingly, this sandbox sent
15
the most fingerprints over a time period of 20 minutes, meaning it restarted
the app over and over, probably trying to maximize code coverage. Mobile
Sandbox also applies some machine learning algorithm to the results and sent
data. However, it shares some of the flaws all Android emulator-based sandboxes
have, like the missing ping and a static IP address.
SandDroid
SandDroid is a Chinese online analysis tool [
34
]. This sandbox performs
detailed static analysis which it presents graphically to the user. It is powered
by Androguard and DroidBox, and allows to upload APK files of up to 20 MB.
Besides its rich static analysis, it features a dynamic analysis and also gives a
security rating. The fingerprint of Sand-Finger never reached our web server, so
HTTP is probably sandboxed, too. Requests returned “302:Found”. However
after uploading a non-HTTPS version of Sand-Finger, the fingerprint results
could simply be copied from SandDroid’s traffic capturing. The timestamp
inside the sandbox is wrong. According to the low Android version, SandDroid
probably uses TaintDroid.
TraceDroid
The TraceDroid Analysis Platform (TAP) is the only sandbox that rarely
performs static analysis and does not rate the results [
35
]. Instead, it packs
all results in an archive that can then be downloaded from a web service. The
data is useful, especially the control flow graphs. Regarding the fingerprint, it
is interesting that many log files are stored on the SD card. TraceDroid uses
Monkey Executor to generate random user input. The sandbox is easily traceable
because the IP of the sandbox is equal to the static IP of the TraceDroid web
server located at the VU University Amsterdam.
Trend Micro
Another anti-virus scanner that is used by the Mobile Sandbox and sends
fingerprints to our web server is Trend Micro. The fingerprinted sandbox of
16
Trend Micro has the model string TrendMicroMarsAndroidSandbox and its IP
address is located in Japan (where this company is located). It is a sandbox
running a rather recent Android version, namely 4.1., and probably builds upon
DroidBox. Interestingly, the user name is always “mars” and the SD card
contains a PCAP-file and a file named mars conf.txt. Hence, this environment
can be fingerprinted with static strings.
Unknown ARM and x86
We received fingerprints from more AV engines that we could not match to a
specific operator yet. These fingerprints were watermarked with the Sand-Finger
samples that we submitted to the Mobile Sandbox. The Mobile Sandbox submits
samples to more AV companies, including ClamAV,Kaspersky,Sophos, and
others. The yet unknown ARM-based sandbox seems to be built on DroidBox
and its IP address is located in China, pointing to a Chinese AV company. The
sandbox is rooted but has no super user APK installed. While the fingerprint
states the device is a google/occam/mako, thus mimicking the Google Nexus 4,
the Android version in reality is 4.1.2 and is based on the Goldfish hardware. The
unknown x86 sandbox is similar and also mimics a Nexus device with a higher
fingerprint than the platform version. It claims to be the google/soju/crespo,
i.e., a Nexus S, but only runs Android 4.0.4. Moreover, it claims to be an
“armeabi-v7a” architecture although it is actually x86-based. This implies that,
just as Joe Sandbox Mobile, this emulator has the ARM-to-x86 binary translator
libhoudini.so installed. The x86 sandbox runs on a dedicated server in France.
4.3. Summary
All of the fingerprinted sandboxes in our test set have specific elements in
common: (1) None of the analysis environments could send an ICMP ping packet
except Joe Sandbox Mobile. (2) No sandbox has been up for more than ten
minutes except the ForeSafe sandbox, which has been up for 15 minutes. This
implies that no sandbox ran our Sand-Finger sample for more than 15 minutes.
(3) None of the tested sandboxes returned a WiFi connection or stored any WiFi
17
SSIDs. Only one sandbox was virtually connected to a cellphone network, but
real devices can send back their fingerprints only if either a WiFi or a cellphone
network is connected. (4) Most of the VMs are simply based on the emulator
hardware Goldfish. (5) The timestamp was often incorrect. The timestamp,
however, is clearly not the best sandbox detection method.
Considering current versions of the sandboxes we tested, we can conclude that
using the uptime of a system in combination with its hardware string and a list
of connected networks is a reliable indicator to differentiate between virtualized
environments and real devices. The false positive rate of this combination would
already be minimal, possibly concerning only devices that were recently switched
on. Some of these indicators can be fixed by sandboxes in the future, but in
turn the indicators used by split-personality malware can be changed, as well.
Hence, creating an Android sandbox that cannot be distinguished easily from
real devices remains an interesting research topic for malware analysis.
5. Defeating Analysis Environments
Insights from our fingerprinting tool Sand-Finger can be deployed to build
split-personality malware that behaves maliciously on real devices, but gets
classified as benign in analysis environments. To defeat sandboxes in general, we
focus on a technique based on split-personality behavior. These attacks divide
malware samples into a benign and malicious part, such that the malicious part
is hidden from analysis by packing, encrypting or outsourcing of the code. As a
consequence, the malware sample outwardly appears benign and hence, misleads
the AV engine. Only when running on real devices, does split-personality malware
show its true behavior. These methods are especially effective, as they are hiding
from both static and dynamic analysis. Malicious code segments are split in
such a way that a sandbox is not aware of critical code.
Note that split-personality attacks are especially successful against restrictive
operating systems like Android. On an unrooted Android phone, it is impossible
for an app to scan another app during runtime, preventing on-device AV scanners
18
from applying heuristics after a sample has unpacked itself. Contrary to that,
proper virus scanners for Windows desktop PCs cannot be tricked as easily, as
they run in a privileged mode and monitor the address space of third party
apps at runtime. Indeed, many apps exist for Android that suggest effective AV
scanning, but technically these apps can only scan a sample statically before
runtime.
5.1. Collusion Attacks defeat VirusTotal Scanners
Automated analysis usually consists of a static and a dynamic part. In the
static phase, sandboxes first identify granted permissions in the manifest file,
scan through constants, embedded URLs, and possibly library calls. Afterwards,
they execute the code, oftentimes tapping on the screen more or less randomly,
and finally rate the sample. Commonly, Android permissions play a central role
in this rating. By doing so, sandboxes can classify ordinary apps that come as
single APK files. Malware, however, can also be split into two complementary
apps.
Due to the way Android’s permission system works, an app can access
resources with the help of another app, in such a way that it does not need
to have all permissions declared in its own manifest. Those attacks are called
collusion attacks [
36
]: Assuming the app with the desired permissions has already
been trusted by the user on installation, another app can send the requests to
perform a certain action by IPC mechanisms, like sockets or even covert channels.
On a system as complex as Android, many covert channels can be found, for
example shared settings. Even using the volume as a 1-bit channel has been
proposed [
37
]. However, social engineering techniques have to be used to get
a user to download the complementary app. But it is a common practice to
provide additional app packages, like game levels, plug-ins, and more.
The collusion attack is effective against automated static and dynamic analysis
alike, because automated analysis looks at only one app at a time. Also note
that the collusion attack is currently the only method, besides root exploits, that
can trick the analysis of permissions. To evaluate this approach, we used the
19
open source malware AndroRAT (Android Remote Administration Tool). This
software connects to a server which then controls the whole Android system
remotely. Many AV scanners detect the installation of AndroRAT as malware,
namely 15/50 AV solutions on VirusTotal. Microsoft’s Windows Defender even
stopped the compilation of AndroRAT from source.
As part of this work, we created three variants of AndroRAT, called AndroRAT-
Welcome,AndroRAT-Pandora and AndroRAT-Split. AndroRAT-Welcome simply
includes an anti-automation screen that must be touched by the user before
the app continues. AndroRAT-Pandora is a obfuscated version of AndroRAT
that we created with the transformation framework PANDORA [
8
] to confuse
static analysis. AndroRAT-Split finally splits the app into two and performs a
collusion attack as described above. Technically, we kept the main service class
in one APK and added the activity classes to a second app. Thanks to Android’s
IPC mechanisms, this approach was straightforward to implement.
While the detection rate of the original AndroRAT sample on VirusTotal
is 15/50, the detection rate of AndroRAT-Welcome declined to 9/15. Thus,
six AV engines were either not able to bypass the welcome screen, or detecting
the original version was based on static signatures, like a hash sum of the
AndroRAT APK. While these detection rates are already fairly low, our goal was
to decrease them further. The detection rate of AndroRAT-Pandora declined to
4/15, probably defeating static analysis systems in the first place. The detection
rate of AndroRAT-Split finally declined to 0/15 for both of the separated apps,
since none of them shows malicious behavior without the other.
We also tested the collusion attack against our test set of dynamic sandboxes.
When analyzing with TraceDroid, a sandbox without static analysis, none of
the malicious methods was detected. On sandboxes that combine static and
dynamic analysis, however, only the activity app was rated as benign. This is
plausible, as without the service class, the app has no harmful code. The service
app was generally considered less harmful but still precarious. Hence, we could
not surpass all sandboxes with a simple collusion attack. More advanced variants
of AndroRAT-Split would be required that either split up the permissions into
20
more apps, or combine the collusion attack with fingerprinting and dynamic
code loading, as described before.
5.2. Dynamic Code Loading defeats the Google Bouncer
In this section, we describe how split-personality malware can be combined
with dynamic code loading to enter the Google Play Store. To enter the Play
Store, an Android app must pass the Google Bouncer, which is the most promi-
nent example of dynamic analysis for Android today.
With respect to dynamic code loading, Android imposes minimal restrictions
inside an app (whereas Apple, for example, tries to restrict dynamic code loading).
From an AV engine’s point of view, the problem is that dynamic code loading is
a common practice in many benign Android apps, too. Just like on Windows,
Android apps are free to load whatever they please at runtime. On Windows,
however, effective malware scanners can be installed that monitor other apps at
runtime.
As opposed to this, on Android every app runs its own Dalvik VM instance
that cannot be accessed by other apps. While this is, on the one hand, secure
because apps are not allowed to alter other apps in a malicious way, apps can do
anything they have permissions for. Once an Android malware sample defeats
the Google Bouncer and enters the Play Store, e.g., based on split-personality
behavior, there is no effective authority that monitors an app while running on
real devices.
The possibilities for virus scanners to counteract threats on Android are
therefore astonishingly small. Static analysis can understand every instruction
but doesn’t see the important parts as they are loaded later and dynamic analysis
could easily analyze the code loaded at runtime, except it is not loaded at all in
this case [
16
,
15
,
17
]. On Windows, this kind of attack would prove impossible
thanks to anti-malware solutions that monitor running apps. It’s clear to see
the security against dynamic code loading is currently insufficient on Android.
Dynamic code can be downloaded from the Internet, copied from SD-card, loaded
from other app’s packages or even ship hidden in the own app’s resources [
14
,
7
].
21
Apart from static analysis, that will not be able to find the dynamically loaded
code, loading the code only while not being watched will also protect the malware
from dynamic analysis. So combining dynamic code loading with one of the split
personality methods discussed in Sect. 4 results in a very well hidden malware.
So even a thoroughly malware-scanned app can do anything as the scanner stops
watching [12].
From a technical point of view, dynamically loaded code in Android can
either be a DEX file or a native library. It can be downloaded from the Internet,
copied from SD card, loaded from other apps packages or be shipped encrypted
or packed in the apps resources [
7
,
14
]. Apart from static analysis, which is
naturally not able to find dynamically loaded code, dynamic analysis can be
defeated when it is combined with fingerprinting techniques. Even a thoroughly
scanned app can do anything after the scanner stops watching [12].
As we revealed during our analysis in Sect. 4, it is currently impossible to
fingerprint the Google Bouncer because it blocks any Internet traffic to custom
servers (or does not execute uploaded APKs at all). This, however, has been
different in the past. Oberheide and Miller were able to fingerprint the Google
Bouncer in 2012 [
11
]. For example, they found out that the Google Bouncer is
based on the Android emulator as it creates the file /sys/qemu trace during boot.
As part of this work, the Google Bouncer was revisited. Since most devices are
regularly connected to the Internet today, the missing Internet connection itself
can be used as a fingerprint. That is, malware can disable malicious code on a
missing Internet connection to evade dynamic analysis of the Google Bouncer.
While cutting the Internet connection on a sandbox hinders fingerprinting, there
is no way of executing malware that relies on connectivity to a command and
control server.
For our tests, we equipped an otherwise benign app with a root exploit that
is executed as soon as a certain command is received from a C&C server which
is regularly contacted. For our first test, we included the root exploit as a plain
native ARM binary into the app. Containing an executable root exploit, the
app should be rejected based on static analyses of the Google Bouncer and
22
not become available for download. As we expected, this variant was indeed
classified as malicious and got rejected by the Bouncer soon after upload. As
a consequence, the author’s Google Play account was banned. Unfortunately,
Google managed to link the banned account to all test accounts, although those
were issued with fake identities, such that all of four developer accounts were
closed at once. It is worth mentioning that the rejected sample was neither
blacklisted by Verify Apps nor uploaded to the Google-owned VirusTotal, despite
its being banned from the store.
Surprisingly, to confuse static analyses by Google’s Bouncer, it turned out to
be sufficient to zip the root exploit. Contrary to our own expectations, it was
neither necessary to encrypt the root exploit, nor to dynamically download it
from the C&C server. The Google Bouncer never contacted our server during its
analysis, but the malicious APK became available for download on the Play Store
a few hours after submission. To verify our approach, we downloaded the binary
to a vulnerable Samsung Galaxy S3 device and ran the exploit. Note that, as our
server always declined execution requests from unknown clients, there was no
danger for Android users at any time. Only when the server permitted execution
was the binary unpacked and executed. Even if the bouncer initially didn’t run
the binary but waited for the app to get a certain number of downloads, it had
no way to detect our malware, so we took our sample from the Play Store within
24h after all our tests had been applied, and we did not count any download of
the app other than ours.
To sum up, malicious apps hiding a root exploit can easily surpass widespread
security checks for Android. Malware without root exploits, however, would still
require declaring its permissions in the manifest file. However it can pair this
method with a collusion attack.
6. Dynamic Script Loading Attacks
In an environment where loading binaries is prohibited, malware authors can
still find an attack surface for dynamic code loading. Shipping a full-fledged
23
scripting environment with an app is one of them. This method has already
been applied by Windows malware like “Flame” [
20
]. Scripts can execute certain
functions inside the binary and considering the spread of scripting environments
in benign apps, flagging an app simply as malicious due to its use of a scripting
host is no solution. Above that, unknown interpreters might be hard to track
for automated analysis.
6.1. Taming Dynamic Code Loading
Today the Android OS lacks a way to prevent attacks based on dynamic
code loading. However, several countermeasures of this type have been proposed
in the past, discussed in this section, that could make more advanced attacks
necessary in the future.
Fedler et al. also proposed several methods to secure the Android platform
against the execution of malicious native code, especially root exploits [
14
]: An
executable bit for Javas System.load() and System.loadLibrary() methods, a
permission based approach where execution is granted per app, and the ability
to load libraries only from specified paths. Above that, Fedler et al. proposed
an anti-virus API for Android that allows certified scanners to monitor the file
system and third party apps on demand [
13
]. For example, live change monitoring
on the file system becomes possible similar to AV solutions on Windows. This
API, if it were included in standard Android, could also help to mitigate the
risks of dynamic code loading.
A more sophisticated approach towards taming dynamic code loading, re-
garding both native code and bytecode, has been presented by Poeplau et al. [
12
].
Poeplau et al. suggest disallowing the execution of unsigned code entirely. They
propose a patch to the Android system that allows multiple signing authorities
and hence fits the Android ecosystem such that other companies, apart from
Google, can still offer software in their own stores. If code is always required to be
signed, so that only trusted code can be executed, on-device AV scanners would
no longer be needed. Revising and signing all code of an app store, however, is
obviously a bottleneck.
24
1//Get an object from the map
2//call function
3//store return to the map again
4Object obj = heap.get(objStr);
5cls = obj.getClass() ;
6Method method = cls.getMethod(name, paramTypes);
7heap.put(var,method.invoke(obj, paramInstances));
Listing 1: Get object from “heap” and call function via reflection
6.2. Proof-of-Concept Implementation
In this section we provide a proof of concept, building on top of the V8
JavaScript Engine that is preinstalled on all Android systems. While it is
possible to use V8 binaries directly, we did not want to load a native library,
such that we took a detour to get to the interpreter right out of Java code. To
this end, we leveraged Android’s WebView, an UI element on Android that can
show webpages and run JavaScript code. As it is possible to inject callbacks into
the JavaScript code, JavaScript can call back to our app’s code. This use of the
WebView has indeed been noted before [
38
,
18
,
39
], but as of now, no usages as
obfuscation method have been observed. To not be bound to preprogrammed
code, we also show the use of reflection to call Anroid APIs out of JavaScript.
For our needs, the WebView had to be set up invisible and with JavaScript
enabled. Scripts are disabled by default for security reasons, to keep developers
from accidentally introduce scripting attacks to their apps. We then set our class
as callback to be able to call it from our JavaScript code. The methods we want
to be able to call from JavaScript need to be annotated with @JavascriptInterface.
To make this technique a proper obfuscation method, we added fingerprinting
from Sect. 4 to the app and made the app to only execute when being run on an
actual phone [
15
,
16
,
17
]. After the setup, we could successfully call pre-prepared
methods in our classes from a webpage that has been loaded from the web.
Of course our problem is, especially if we want to use the Android API, that
we need to use Java objects but cannot return or construct them inside the
25
WebView. Note that this problem might not occur using alternative interpreters
like Rhino
1
. To be able to use the Java VM anyway, we wrapped the Java
reflection API once more: instead of returning objects, we store them into
a hashmap that maps a string to an object. The key can be specified from
JavaScript and can then be used as reference for following function calls. Thanks
to this virtual heap, we can call any Java methods with any parameters by
simply storing them in the map and using the strings to specify them. The
use of reflection can be seen in Listing 1. Our wrapper function takes care of
mapping the given strings to objects. However, as we need to be able to read
actual return values for conditionals and loops, we implemented an additional
method that returns a string representation of the value to a given key inside the
hashmap. To pass initial parameters, we added functions to store strings into
the hashmap. The high-level overview over the interactions is shown in Fig. 3.
With only these three methods, calling a function using reflection, storing a
string and getting a string back, a large part of the Android functions can be
used. For convenience, we added methods to call constructors, get field values,
pass integer values and arrays back and forth and free unused hashmap entries
but all those could have been implemented in JavaScript as well. Additionally,
before loading any JavaScript, we filled the hashmap with one value, namely the
activity’s context. A context in Android is needed for many API-calls in Android.
For it we chose the key
ctx
. A lot of work has been performed on Java reflection,
but by now it is still only possible to analyze reflection by combining static and
dynamic analysis [
40
,
41
]. Using this setup, we were able to periodically send
SMS containing the last known GPS position out of JavaScript using the code
seen in Listing 3 inside JavaScript’s
setInterval()
function. As comparison,
the pure Java code to our JavaScript example is given in Listing 2. As to our
expectation, none of the tested sandbox tools could spot this attack.
1urlhttps://developer.mozilla.org/de/docs/Rhino
26
1SmsManager.getDefault() .sendTextMessage(phoneNumber, null, ”EvlSms:
” + location.substring(0, location.length() 40), null,null);
Listing 2: Java code to send an SMS
1// Put a string into the hashmap
2app.createString( ’phoneNumber’, NUMBER);
3app.createString( ’ text’ , SMSTEXT);
4// call static method and store return as smsManager
5app.callFunction(smsManager’,
6android.telephony.SmsManager’,
7’ getDefault’ , [] ’ );
8// a class name as param instead of a key stands for null
9app.callMethod(’void’, smsManager’, ’sendTextMessage’,
’[”phoneNumber”, ”java.lang.String”, ”text”,
”android.app.PendingIntent”, ”android.app.PendingIntent”]’));
Listing 3: JavaScript code to send an SMS via reflection.
7. Countermeasures
As seen in the previous section, prohibiting dynamic code loading on Android
may not solve the problem of split-personality malware but yield an arms race
where the attacker’s next step would be to use dynamic script loading. In this
section, we now discuss other countermeasures that have been proposed against
the problem of split-personality behavior in the literature.
7.1. Machine Learning
Instead of playing the ongoing game of cat-and-mouse between AV scanners
and malware authors, many researchers propose using machine learning and
anomaly detection algorithms [
42
,
43
]. In theory, a well-trained machine learning
algorithm not only identifies known malware samples, or those matching certain
heuristics, but also yields to correct results for yet unknown malware. As an
advantage, machine learning algorithms evolve over time and can learn from a
27
few known malware samples that were analyzed manually. With a growing set
of analyzed apps, the detection rate of machine learning becomes more accurate.
Hence, this approach is particularly suitable for sandboxes, and some of the
sandboxes we tested already started to apply machine learning algorithms, like
the Mobile Sandbox.
7.2. Improving Sandboxes
Many ideas that are used in both mobile malware and mobile sandboxes,
have been found on desktop PCs before. For the case of desktop PCs, Kirat et al.
proposed to run malware analysis on bare-metal [
22
], i.e., on real devices rather
than virtualized environments. To our knowledge, a similar approach has not
been proposed or implemented for Android yet. From the hardware perspective,
it would be feasible though, building upon the possibility to root devices and by
communicating with an analysis PC via Android’s debugging bridge.
Additionally, also for desktop PCs, Sun et al. proposed a system to detect
virtualization resistant behavior in malware [
44
]. To this end, they measure
the distance in behavior between running an app on real devices and running
it inside VMs. A similar approach is taken by Lindorfer et al. [
45
]. Their tool
figures out evasive malware by running in different environments. However, this
tool does not compare the run to a real device, but compares runs on different
VMs. In the future, similar approaches could also be taken on Android to tackle
split-personality malware and fingerprinting.
As opposed to the aforementioned improvements to run Android malware on
bare-metal, another future research project could involve the implementation
of a transparent malware sandbox in software. As seen in Sect. 4, today it is
rather simple to fingerprint available Android sandboxes. The implementation
of a solution that is almost indistinguishable from real hardware remains an
interesting future task. As a starting point, the fingerprinting tool Sand-Finger
could be used as an opponent that must be defeated. Obvious flaws such as the
missing ability to ping hosts on an unaltered QEMU emulator must be addressed.
In addition, personal phone data could be faked, like sensor data, photos and
28
contacts.
7.3. Secure IPC Mechanisms for Android
As shown by collusion attacks in Sect. 5.1, IPC mechanisms on Android
can be used to unite the permissions of two distinct apps. An app can use the
permissions of another app without a user’s approval. To tackle this problem,
XManDroid [
37
] has been proposed by Bugiel et al. to patch the Android
platform against the risk of privilege escalations [
46
]. XManDroid claims to
also stop collusion attacks that communicate via channels other than Android’s
standard IPC mechanism.
8. Conclusions
On-device AV scanners for Android have little possibilities of counteracting
threats. Once a malware sample is installed, there are no possibilities to analyze it
at runtime, i.e., no possibilities to detect its malicious behavior. As a consequence,
the Google Bouncer was introduced to reject suspicious apps and to keep the
Play Store clean. As shown by our work, however, the Google Bouncer can
be surpassed by dynamic code loading attacks today. Even if dynamic code
loading would be prohibited or restricted on Android in future, as proposed in
the literature [
12
], split-personality malware cannot be stopped as shown by
dynamic script loading attacks.
As part of this work, ten unique sandboxes have been fingerprinted using
a custom-built fingerprinter called Sand-Finger. Moreover, we argued that
results of our fingerprinter can be applied to Android malware to perform split-
personality attacks that defeat automated analysis frameworks. To substantiate
our approach, we demonstrated that dynamic code loading can be used to bypass
the Google Bouncer, and successfully uploaded an Android root exploit to the
Play Store. We also revised the effectiveness of so-called collusion attacks based
on the AV detection rates of VirusTotal.
Furthermore, we have studied the widespread of dynamic code loading among
22,032 benign and 14,885 malicous apps in the wild. The results show that
29
benign apps use these techniques less frequently than malicious ones, among
which 36.4% were found to dynamically load and execute code by means of class
loaders and the Runtime.exec() method.
To conclude, neither automated static nor dynamic analysis of Android APKs
can be considered entirely secure today and foreseeable countermeasures are
just the next step in an ongoing game of cat-and-mouse. Static analysis has no
possibility of analyzing code or scripts which an app dynamically downloads or
unpacks at runtime before execution. Dynamic analysis, on the other hand, can
easily analyze code and scripts loaded at runtime, but it cannot analyze traces
which are not executed inside an analysis environment. Hence, from an attackers
point of view, the trick is to behave differently in analysis environments than on
real devices.
Acknowledgments
The research leading to these results was supported by the Bavarian State
Ministry of Education, Science and the Arts as part of the FORSEC research
association. We want to thank Johannes G¨otzfried, Andreas Kurtz, Mykola
Protsenko and Michael Spreitzenbarth for proofreading this article and giving us
valuable hints for improving it. A special thanks also goes to Sebastian Poeplau
and his co-authors for early access to their related paper [12].
References
[1]
Gartner, Gartner Says Smartphone Sales Accounted for 55 Percent of Overall
Mobile Phone Sales in Third Quarter of 2013 (Nov. 2013).
URL http://www.gartner.com/newsroom/id/2623415
[2]
D. Balzarotti, M. Cova, C. Karlberger, C. Kruegel, E. Kirda, G. Vigna,
Efficient detection of split personalities in malware, in: In Proceedings of
the Symposium on Network and Distributed System Security (NDSS), San
Diego, CA, 2010.
30
[3]
M. Spreitzenbarth, F. C. Freiling, F. Echtler, T. Schreck, J. Hoffmann,
Mobile-sandbox: Having a deeper look into Android applications, in:
Proceedings of the 28th Annual ACM Symposium on Applied Com-
puting, SAC ’13, ACM, New York, NY, USA, 2013, pp. 1808–1815.
doi:10.1145/2480362.2480701.
[4]
X. Chen, J. Andersen, Z. Mao, M. Bailey, J. Nazario, Towards an un-
derstanding of anti-virtualization and anti-debugging behavior in modern
malware, in: Dependable Systems and Networks With FTCS and DCC,
2008. DSN 2008. IEEE International Conference on, 2008, pp. 177–186.
doi:10.1109/DSN.2008.4630086.
[5]
A. Shabtai, Y. Fledel, U. Kanonov, Y. Elovici, S. Dolev, C. Glezer, Google
Android: A comprehensive security assessment, IEEE Security and Privacy
8 (2) (2010) 35–44. doi:10.1109/MSP.2010.2.
[6]
V. Rastogi, Y. Chen, X. Jiang, Droidchameleon: Evaluating Android anti-
malware against transformation attacks, in: Proceedings of the 8th ACM
SIGSAC Symposium on Information, Computer and Communications Se-
curity, ASIA CCS ’13, ACM, New York, NY, USA, 2013, pp. 329–334.
doi:10.1145/2484313.2484355.
[7]
R. Fedler, J. Sch¨utte, M. Kulicke, On the effectiveness of malware protection
on Android, Tech. rep., Fraunhofer AISEC, Berlin (2013).
[8]
Mykola Protsenko, Tilo M¨uller, PANDORA Applies Non-Deterministic
Obfuscation Randomly to Android, in: Fernando C. Osorio (Ed.), 8th
International Conference on Malicious and Unwanted Software, 2013.
[9] H. Lockheimer, Android and Security (Feb. 2012).
URL
googlemobile.blogspot.de/2012/02/android-and-security.htm
[10]
N. J. Percoco, S. Schulte, Adventures in Bouncerland: Failures of Automated
Malware Detection with in Mobile Application Markets, in: Black Hat USA
’12, Trustwave SpiderLabs, 2012.
31
[11]
J. Oberheide, C. Miller, Dissecting the Android Bouncer, Presented at
SummerCon 2012, Brooklyn, NY, 2012.
URL https://jon.oberheide.org/files/summercon12-bouncer.pdf
[12]
S. Poeplau, Y. Fratantonio, A. Bianchi, C. Kruegel, G. Vigna, Execute
This! Analyzing Unsafe and Malicious Dynamic Code Loading in Android
Applications, in: Proceedings of the ISOC Network and Distributed System
Security Symposium (NDSS), San Diego, CA, 2014.
[13]
R. Fedler, M. Kulicke, J. Sch¨utte, An antivirus API for Android mal-
ware recognition, in: Malicious and Unwanted Software: ”The Ameri-
cas” (MALWARE), 2013 8th International Conference on, 2013, pp. 77–84.
doi:10.1109/MALWARE.2013.6703688.
[14]
R. Fedler, M. Kulicke, J. Schuette, Native code execution control for attack
mitigation on Android, in: Proceedings of the Third ACM Workshop on
Security and Privacy in Smartphones & Mobile Devices, SPSM ’13,
ACM, New York, NY, USA, 2013, pp. 15–20. doi:10.1145/2516760.2516765.
[15]
D. Maier, T. M¨uller, M. Protsenko, Divide-and-Conquer: Why Android Mal-
ware cannot be stopped, in: S. Research (Ed.), 9th International Conference
on Availability, Reliability and Security, 2014.
[16]
T. Petsas, G. Voyatzis, E. Athanasopoulos, M. Polychronakis, S. Ioan-
nidis, Rage against the virtual machine: Hindering dynamic analysis of
android malware, in: Proceedings of the Seventh European Workshop on
System Security, EuroSec ’14, ACM, New York, NY, USA, 2014, pp. 5:1–5:6.
doi:10.1145/2592791.2592796.
[17]
T. Vidas, N. Christin, Evading android runtime analysis via sandbox detec-
tion, in: Proceedings of the 9th ACM Symposium on Information, Computer
and Communications Security, ASIA CCS ’14, ACM, New York, NY, USA,
2014, pp. 447–458. doi:10.1145/2590296.2590325.
32
[18]
T. Luo, H. Hao, W. Du, Y. Wang, H. Yin, Attacks on WebView in the
Android system, in: Proceedings of the 27th Annual Computer Security
Applications Conference, ACSAC ’11, ACM, New York, NY, USA, 2011,
pp. 343–352. doi:10.1145/2076732.2076781.
[19]
T. Raffetseder, C. Kruegel, E. Kirda, Detecting system emulators, in:
Proceedings of the 10th International Conference on Information Security,
ISC’07, Springer-Verlag, Berlin, Heidelberg, 2007, pp. 1–18.
[20]
B. Bencs´ath, G. P´ek, L. Butty´an, M. Felegyhazi, sKyWIper (a.k.a. flame
a.k.a. flamer): A complex malware for targeted attacks, Tech. rep., In
collaboration with the sKyWIper Analysis Team (2012).
[21]
A. Moser, C. Kruegel, E. Kirda, Limits of static analysis for malware detec-
tion, in: Computer Security Applications Conference, 2007. ACSAC 2007.
Twenty-Third Annual, 2007, pp. 421–430. doi:10.1109/ACSAC.2007.21.
[22]
D. Kirat, G. Vigna, C. Kruegel, Barebox: Efficient malware analysis on bare-
metal, in: Proceedings of the 27th Annual Computer Security Applications
Conference, ACSAC ’11, ACM, New York, NY, USA, 2011, pp. 403–412.
doi:10.1145/2076732.2076790.
[23]
C. Willems, T. Holz, F. C. Freiling, Toward automated dynamic malware
analysis using cwsandbox, IEEE Security and Privacy 5 (2) (2007) 32–39.
doi:10.1109/MSP.2007.45.
[24]
U. Bayer, I. Habibi, D. Balzarotti, E. Kirda, C. Kruegel, A view on current
malware behaviors, in: Proceedings of the 2Nd USENIX Conference on
Large-scale Exploits and Emergent Threats: Botnets, Spyware, Worms, and
More, LEET’09, USENIX Association, Berkeley, CA, USA, 2009, pp. 8–8.
[25]
M. Becher, F. C. Freiling, Towards dynamic malware analysis to increase
mobile device security, in: A. Alkassar, J. H. Siekmann (Eds.), Sicherheit,
Vol. 128 of LNI, GI, pp. 423–433.
33
[26]
M. Becher, R. Hund, Kernel-level interception and applications on mobile
devices, Tech. rep., University of Mannheim (may 2008).
[27]
T. Bl¨asing, L. Batyuk, A.-D. Schmidt, S. Camtepe, S. Albayrak, An Android
application sandbox system for suspicious software detection, in: Malicious
and Unwanted Software (MALWARE), 2010 5th International Conference
on, 2010, pp. 55–62. doi:10.1109/MALWARE.2010.5665792.
[28]
R. Vall´ee-Rai, P. Co, E. Gagnon, L. Hendren, P. Lam, V. Sundaresan,
Soot - A Java Bytecode Optimization Framework, in: Proceedings of the
Conference of the Centre for Advanced Studies on Collaborative Research,
CASCON ’99, IBM Press.
[29]
M. Spreitzenbarth, Dissecting the Droid: Forensic Analysis of Android and
its malicious Applications, Ph.D. thesis, Friedrich-Alexander-Universit¨at
Erlangen-N¨urnberg (2013).
[30]
M. Lindorfer, Andrubis: A Tool for Analyzing Unknown Android Applica-
tions (Jun. 2012).
[31]
Anthony Desnos, Androguard: Reverse engineering, Malware and goodware
analysis of Android applications.
URL code.google.com/p/androguard/
[32]
W. Enck, P. Gilbert, B.-G. Chun, L. P. Cox, J. Jung, P. McDaniel, A. N.
Sheth, Taintdroid: An information-flow tracking system for realtime privacy
monitoring on smartphones, in: Proceedings of the 9th USENIX Conference
on Operating Systems Design and Implementation, OSDI’10, USENIX
Association, Berkeley, CA, USA, 2010, pp. 1–6.
[33] ForSafe, Forsafe mobile security, Whitepaper (2013).
URL http://www.foresafe.com/ForeSafe_WhitePaper.pdf
[34]
Botnet Research Team, SandDroid: An APK Analysis Sandbox, Xi’an
Jiaotong University (2014).
URL http://sanddroid.xjtu.edu.cn
34
[35]
V. Van der Veen, Dynamic Analysis of Android Malware, Master Thesis,
VU University Amsterdam (Aug. 2013).
URL http://tracedroid.few.vu.nl/thesis.pdf
[36]
C. Marforio, H. Ritzdorf, A. Francillon, S. Capkun, Analysis of the Com-
munication Between Colluding Applications on Modern Smartphones, in:
Proceedings of the 28th Annual Computer Security Applications Conference,
ACSAC ’12, ACM, New York, NY, USA, 2012, pp. 51–60.
[37]
S. Bugiel, L. Davi, A. Dmitrienko, T. Fischer, A.-R. Sadeghi, B. Shastry,
Towards Taming Privilege-Escalation Attacks on Android, in: 19th Annual
Network & Distributed System Security Symposium, NDSS ’12, 2012.
[38]
E. Chin, D. Wagner, Bifocals: Analyzing webview vulnerabilities in android
applications, in: Proceedings of the 14th International Workshop (WISA),
2013, pp. 138–159.
[39] M. Neugschwandtner, M. Lindorfer, C. Platzer, A view to a kill: Webview
exploitation, in: Presented as part of the 6th USENIX Workshop on Large-
Scale Exploits and Emergent Threats, USENIX, Berkeley, CA, 2013.
[40]
E. Bodden, A. Sewe, J. Sinschek, H. Oueslati, M. Mezini, Taming reflec-
tion: Aiding static analysis in the presence of reflection and custom class
loaders, in: Proceedings of the 33rd International Conference on Software
Engineering, ICSE ’11, ACM, New York, NY, USA, 2011, pp. 241–250.
doi:10.1145/1985793.1985827.
[41]
H. Hao, V. Singh, W. Du, On the effectiveness of api-level access control
using bytecode rewriting in android, in: Proceedings of the 8th ACM
SIGSAC Symposium on Information, Computer and Communications
Security, ASIA CCS ’13, ACM, New York, NY, USA, 2013, pp. 25–36.
doi:10.1145/2484313.2484317.
[42]
A. Shabtai, U. Kanonov, Y. Elovici, C. Glezer, Y. Weiss, “Andromaly”: a
behavioral malware detection framework for Android devices, Journal of
35
Intelligent Information Systems 38 (1) (2012) 161–190. doi:10.1007/s10844-
010-0148-x.
[43]
D. Arp, M. Spreitzenbarth, M. Hubner, H. Gascon, K. Rieck, DREBIN:
Effective and Explainable Detection of Android Malware in Your Pocket,
in: Proceedings of the ISOC Network and Distributed System Security
Symposium (NDSS), San Diego, CA, 2014.
[44]
M.-K. Sun, M.-J. Lin, M. Chang, C.-S. Laih, H.-T. Lin, Malware
virtualization-resistant behavior detection., in: ICPADS, IEEE, 2011, pp.
912–917. doi:10.1109/ICPADS.2011.78.
[45]
M. Lindorfer, C. Kolbitsch, P. Milani Comparetti, Detecting environment-
sensitive malware, in: Proceedings of the 14th International Conference on
Recent Advances in Intrusion Detection, RAID’11, Springer-Verlag, Berlin,
Heidelberg, 2011, pp. 338–357.
[46]
L. Davi, A. Dmitrienko, A.-R. Sadeghi, M. Winandy, Privilege escalation
attacks on Android, in: Proceedings of the 13th International Conference
on Information Security, ISC’10, Springer-Verlag, Berlin, Heidelberg, 2011,
pp. 346–360.
36
Sandbox
Android Version
CPU Architecture
Resembles Hardware
Successful Ping
Correct Timestamp
Hardware not Goldfish
WiFi Connected
Cell Info
UpTime >10 min.
Consistent
Untraceable
Google Nexus 5 4.4.2 armv7l XXXXXXXX-
SDK Emulator (x86) 4.4.2 i686 XXXX X XXX-
1 Andrubis 2.3.4 armv5tejl XXXXXXXX-
2 BitDefender 4.3 i686 XXXX X XXXX
3 ForeSafe 2.3.4 armv5tejl X X XXXXX X X
4 Joe Sandbox Mobile 4.2.1 i686 XXXXXXXX X
5 Mobile Sandbox 4.1.2 armv7l XXX X XXXX-
6 SandDroid 2.3.4 armv5tejl XXXXXXXX-
7 TraceDroid 2.3.4 armv5tejl X X XXXXXXX
8 Trend Micro 4.1.1 armv7l XXXXXXXXX
9 Unknown ARM 4.1.2 armv7l XXXXXXXXX
10 Unknown x86 4.0.4 i686 XXXXXXXXX
Figure 2: Fingerprinting ten Android sandboxes with Sand-Finger. The eleven most revealing
values (out of over 100) are listed here as examples.
37
Hashmap (“Heap”) Android Eventloop JavaScript Loop
Store string
Store string
Call function using reflection
request parameter
values
Call actual function
Store return value
Success/Error
Request value
Request value
Value
Value’s string representation
Figure 3: Interaction between JavaScript and the app.
38
... Scientists studied the possibility of detecting the running environment fingerprints to differentiate between an emulator and a physical device (Jing et al., 2014;Maier, Muller & Protsenko, 2014;Maier, Protsenko & Müller, 2015;. Android.obad ...
... As shown in Table 9, many static analysis frameworks examine the robustness of their detection frameworks against REF evasion, such as DroidAPIMiner (Aafer, Du & Yin, 2013), DexHunter (Zhang, Luo & Yin, 2015), SherLockDroid (Apvrille & Apvrille, 2015), Kuhnel (Kuhnel, Smieschek & Meyer, 2015), DroidRA , and AAMO. Likewise, Maier (Maier, Protsenko & Müller, 2015), which uses Dynamic analysis, RiskRanker (Grace et al., 2012), and StaDyna (Zhauniarovich et al., 2015), which use hybrid analysis, study REF evasion detection using dynamic and hybrid analysis based detection techniques. This study spotted six papers that scrutinized the REF evasion using static analysis, only two papers scrutinized REF using dynamic analysis, and two papers scrutinized REF using hybrid analysis. ...
... -DCL Evasion Detection: Some Android malware detection frameworks propose and evaluate their methods to detect DCL evasion, for instance, DroidAPIMiner (Aafer, Du & Yin, 2013), Poeplau (Poeplau et al., 2014), Dexhunter, Maier (Maier, Protsenko & Müller, 2015), RiskRanker (Grace et al., 2012), and StaDyna (Zhauniarovich et al., 2015). However, AndroSimilar (Faruki et al., 2015d) insufficiently evaluates its mechanism against dynamic code loading, reflection, and other transformation techniques, as shown in Table 9. ...
Article
Full-text available
The various application markets are facing an exponential growth of Android malware. Every day, thousands of new Android malware applications emerge. Android malware hackers adopt reverse engineering and repackage benign applications with their malicious code. Therefore, Android applications developers tend to use state-of-the-art obfuscation techniques to mitigate the risk of application plagiarism. The malware authors adopt the obfuscation and transformation techniques to defeat the anti-malware detections, which this paper refers to as evasions. Malware authors use obfuscation techniques to generate new malware variants from the same malicious code. The concern of encountering difficulties in malware reverse engineering motivates researchers to secure the source code of benign Android applications using evasion techniques. This study reviews the state-of-the-art evasion tools and techniques. The study criticizes the existing research gap of detection in the latest Android malware detection frameworks and challenges the classification performance against various evasion techniques. The study concludes the research gaps in evaluating the current Android malware detection framework robustness against state-of-the-art evasion techniques. The study concludes the recent Android malware detection-related issues and lessons learned which require researchers’ attention in the future.
... This can happen for special update vectors (e.g., if the MAC address of the computer is used as an update parameter). Also, the condition does not protect from split personality malware, i.e., software that acts differently depending on which system it is installed upon, a threat different in nature from the one discussed by Maier et al. [93]. ...
Thesis
Full-text available
This thesis focuses on protecting data in use (while being processed). We investigate the requirements of an execution container which makes it possible to eliminate the need for fully trusting a single service operator/provider. The main approach relies on a set of properties of a trustworthy computational model. The foremost purpose is to protect users’ data privacy in data-centric services ranging from local to remote environments. First, we get a closer look at protecting users from being targeted by software (SW) suppliers or attackers exploiting the SW update systems. The main goal is to ensure update consistency and to avoid freeze updates, i.e. to protect users against malicious SW update. We define the conditions for a secure SW update system that shall allow the user to detect malicious updates. Based on those conditions, we depict the Malicious Update Detector (MUD) framework. Similar to the Certificate Transparency (CT), MUD is designed utilizing Merkle Tree (MT) data structure to keep logs of SW versions. The design aims at achieving transparency of the entries of SW versions in the system using a Primary log and a set of attestor logs. Consequently, that intends to maintain update consistency. Nevertheless, such a design might come with performance overhead. Thereupon, a prototype is evaluated to understand the applicability of the MUD framework on Ubuntu APT as a proof-of-concept. Second, the attention is directed towards the need to protect users’ data in remote environments against malicious attackers and/or honest-but-curious operators/providers. We provide a conceptual abstract definition of a trustworthy execution container, which we call Sealed Computation (SC). The SC is defined via a set of properties, if fulfilled then the system can provide trustworthy service in terms of protecting data in use and software confidentiality. However, hard technologies still cannot provide full protection. Thus, a mutual trust establishment procedure is presented. It employs the properties of SC, which are: Sealing, Attestation, Black-box, and Tamper-resistance, along with the Auditor role. The procedure is designed to take out the need to trust a single entity in the system. It is created for an abstract general architecture of cloud-based applications that includes Data providers and consumers, providers for cloud infrastructure, SW application, and SC container, in addition to the auditing party. The attacker assumption restricts the number of parties that may act maliciously simultaneously. Although this part of the thesis is theoretical, it presents strong preliminary guidelines for designing systems considering data privacy at an early stage, Privacy by Design (PbD). Moreover, several commercial technologies exist which can be viewed as implementations of the SC concept, to some extent, such as Intel’s Software Guard Extensions (SGX) and the Hardware Security Modules (HSMs). Then, the notion of SC is extended in a hybrid model to solve the Secure Multiparty Computation resiliently. The designed model is a distributed system of trusted modules (SC modules) in addition to a set of untrusted processes representing the parties who are interested in computing a share function. The model uses reliable broadcast primitives to exchange values among trusted modules. We show that the model fulfills the SMC properties with considerable resilience and availability under the attacker assumptions. Finally, taking the PbD into account, we study a use case that employs the SGX as an approxima- tion(instance) of the SC concept in a smart home application for healthcare service. Unsurprisingly, additional protection comes at a cost. Therefore, a proof-of-concept is implemented to understand the overhead of utilizing SGX in such a data-centric system. We hope that the study provides systems developers with a manifestation of the substantial impact on performance to balance and reduce the security-performance trade-off. That can be achieved by a careful design decision taking the requirements of data operations in their systems into account along with allowing for utilizing the SC properties for stronger protection.
... A malware can run its malicious code when the mobile device is booted or after a certain period of time passed. The dynamic analysis tools in Android generally run for 10 minutes [20]. Hence, attackers can postpone triggering malicious code in a virtual/real device. ...
Conference Paper
Full-text available
With the developments in mobile and wireless technology, mobile devices have become important part of our lives. While Android is the leading operating system in the market share, it is also the most targeted platform by attackers. While there have been many solutions proposed for detection of Android malware in the literature, the family classification of detected malicious applications becomes important, especially where the number of mobile malware variants increases every day in the market. In this study, a solution based on machine learning and hybrid analysis is proposed for the Android malware familial classification problem. An extensive feature set including network-related features and activity bigrams is proposed. The effective static and dynamic analysis features are studied thoroughly and evaluated on Malgenome, Drebin, and UpDroid datasets.
... Aysan and Sen [10] showed that 19.60% of 25,000 applications from three different markets use DCL for the purpose of updating. The research conducted by Maier et al. [11] on a sample of 14,885 malicious and 22,032 benign applications, showed that 36.4% of malicious and 13.1% of benign applications use DCL. Spreitzenbarth et al. [9] have found that 24% of 36,000 applications from alternative sources use native calls. ...
Article
Full-text available
One of the constant challenges faced by the Android community, when it comes to the safety of the end users, is the ability of applications to load code dynamically. This mechanism may be used for both legitimate and malicious purposes. A particular problem is the fact that remote code is not analyzed during the verification process because it doesn’t have to be present in the application package at the publishing time. Previous research has shown that using this concept in an insecure way can cause serious consequences for the user and his device. Solving this problem has proved to be a big challenge that many have tried to address in different ways. This paper deals with the problem of dynamic code loading on Android platform. For the purpose of this paper, an application that demonstrates the abuse of the dynamic code loading concept has been developed and published in the Google Play Store. Also, a proposal of the modified Android ecosystem that should address this problem and improve the security of the whole platform is given.
... This method detects code injection attacks from Android hybrid applications with an accuracy of 97.55%. Maier et al. [24] pointed out that many malicious applications adopt a split-personality behavior to hide malicious codes to avoid detection of antivirus software. Malicious applications use more dynamic code loading techniques than benign ones. ...
Article
We propose a novel hybrid method to analyze the security vulnerabilities in Android applications. Our method combines static analysis, which consists of metadata and data flow analyses with dynamic analysis, which includes dynamic executable scripts and application program interface hooks. Our hybrid method can effectively analyze nine major categories of important security vulnerabilities in Android applications. We design dynamic executable scripts that record and perform manual operations to customize the execution path of the target application. Our dynamic executable scripts can replace most manual operations, simplify the analysis process, and further verify the corresponding security vulnerabilities. We successfully statically analyze 5547 malwares in Drebin and 10151 real-world applications. The average analysis time of each application in Drebin is 4.52 s, whereas it reaches 92.02 s for real-word applications. Our system can detect all the labeled vulnerabilities among 56 labeled applications. Further dynamic verification shows that our static analysis accuracy approximates 95% for real-world applications. Experiments show that our dynamic analysis can effectively detect the vulnerability named input unverified, which is difficult to be detected by other methods. In addition, our dynamic analysis can be extended to detect more types of vulnerabilities.
... To summarize and highlight critical statements from this work, we briefly refer to Both attacker models rely on the premise that malicious apps can pass the boundaries of an app store. This premise has been proven time and time again by applications masquerading as a legitimate version of an official app [206,161,186], by split-personality malware behaving benign during testing [135,150,2], or even relying on collusion attacks where the malicious code and needed permissions are split between two apps [62,114,144]. This premise is valid not only for the Google Play Store but also for other app stores, e.g., the Apple App Store [33,175,121]. ...
Thesis
Full-text available
In this thesis, we investigate different possibilities to protect the Android ecosystem better. We focus on protection mechanisms for application developers, and present modern attacks against sandbox-protected applications and the developer’s intellectual property, ultimately providing enhanced approaches for defense against these attacks. Our defensive approaches range from runtime-shielding measures to analysis-impeding obfuscation mechanisms. First, we take a closer look at communication possibilities of sandboxed applications on Android, namely the UI layer and Android’s inter-process communication. We introduce attacks on applications working through the actors on Android’s UI, starting with overlay windows, accessibility services, input editors, and screen captures. Android’s inter-process communication is the second attack avenue we pursue. It is the primary means of communication for apps to interact with each other despite being sandboxed by the Android system. We show through assessments of the Google Play Store and third-party app stores that attacks on these mechanisms pose a blind-spot in current attack models considered by developers. To provide relief we introduce new protection mechanisms that developers can implement and enhance testing methodologies to consider these attacks in the future. Second, we direct the reader’s attention towards attacks on the developer’s intellectual property. Due to Android’s open-source nature and openly communicated standards, a trend of repackaging popular applications with malicious enhancements and republishing the malicious app has rooted itself in the malware community. To counteract this development, we present an enhanced centroid-based approach at clone detection and improved analysis-impeding obfuscation mechanisms that build on virtualization-based obfuscation. Our obfuscation approach works on Android’s current runtime environment, as well as the previously employed ‘Dalvik virtual machine’, and can be used to obfuscate critical portions of an application’s functionality against prying eyes. To make valid assumptions about the strength of virtualization-based obfuscation, we conduct a de-obfuscation study on the more mature x86/x64 platform, developing a reverse engineering approach for virtualization-obfuscated binaries. We analyzed several hundred thousand Android applications during our research with automated approaches and several thousand apps with manual analysis, always opting for a responsible disclosure process of found vulnerabilities by providing developers with at least three months’ due notice before attempting a publication. The tools presented in this thesis are open-sourced under the MIT license, to help in the inclusion of development projects and their extension or further development. With the insights gained through the research for this thesis, we hope to provide developers with the tools and testing approaches they need to make the Android ecosystem more secure and safe.
Article
Full-text available
Android platform security is an active area of research where malware detection techniques continuously evolve to identify novel malware and improve the timely and accurate detection of existing malware. Adversaries are constantly in charge of employing innovative techniques to avoid or prolong malware detection effectively. Past studies have shown that malware detection systems are susceptible to evasion attacks where adversaries can successfully bypass the existing security defenses and deliver the malware to the target system without being detected. The evolution of escape-resistant systems is an open research problem. This paper presents a detailed taxonomy and evaluation of Android-based malware evasion techniques deployed to circumvent malware detection. The study characterizes such evasion techniques into two broad categories, polymorphism and metamorphism, and analyses techniques used for stealth malware detection based on the malware’s unique characteristics. Furthermore, the article also presents a qualitative and systematic comparison of evasion detection frameworks and their detection methodologies for Android-based malware. Finally, the survey discusses open-ended questions and potential future directions for continued research in mobile malware detection.
Chapter
Extracting evidence pertaining to mobile apps is a key task in mobile device forensics. Since mobile apps can generate more than 19,000 files on a single device, it is time consuming and error prone to manually inspect all the files. Fuzzing tools that programmatically produce interactions with mobile apps are helpful when paired with sandbox environments to study their runtime forensic behavior and summarize patterns of evidentiary data in forensic investigations. However, the ability of fuzzing tools to improve the efficiency of mobile app forensic analyses has not been investigated. This chapter describes AFuzzShield, an Android app shield that protects apps from being exercised by fuzzers. By analyzing the runtime information of mobile app interaction traces, AFuzzShield prevents real-world apps from being exercised by fuzzers and minimizes the overhead on human usage. A statistical model is employed to distinguish between fuzzer and human patterns; this eliminates the need to perform graphical user interface injections and ensures compatibility with apps with touchable/clickable graphical user interfaces. AFuzzShield verifies mobile app program coverage in situations where apps engage anti-fuzzing technologies. Specifically, it was applied to apps in AndroTest, a popular benchmark app dataset for testing fuzzers. The experimental results demonstrate that applying AFuzzShield significantly impacts mobile app program coverage in terms of reduced evidentiary data patterns.
Article
In the past few years, numerous attempts have been made to mitigate evasive Android malware. However, it remains one of the challenges in smartphone security. Evasive malware can dodge dynamic analysis by detecting execution in sandboxes and hiding its malicious behaviors during the investigation. In this work, we present CamoDroid, an open-source and extendable dynamic analysis environment resilient against detection by state-of-the-art evasive Android malware. Our technique mimics data, sensors, user input, static and network features of actual devices and cloaks the existence of the analysis environment. It further improves dynamic analysis and provides a broad view of an application’s behavior by monitoring and logging the dangerous Application Programming Interface (API) calls executed by applications. We implement CamoDroid and assess its resiliency to sandbox detection. We first demonstrate that our sandbox cannot be detected using modern existing academic and commercial applications that can distinguish analysis environments from real devices. We also assess the dependability of CamoDroid against real-world evasive malware and show that it can successfully cloak the existence of the analysis environment to more than 96 percent of evasive Android malware. Moreover, we investigate other popular Android sandboxes and show that they are vulnerable to at least one type of sandbox detection heuristic.
Chapter
At present, most of the new Android devices on the market use the system-as-root architecture. Therefore, the security of personal privacy data under this architecture has become a very concerned issue for the majority of people. As for the problem, this paper proposes to build a safe ecosystem by customizing TF card. When users deal with privacy data, our scheme creates a safe mode based on TF card, and store privacy data in the safe mode. In this way, users can protect their privacy data in the safe modes, even if the mobile is stolen by other people. In order to illustrate the security of the safe mode, this paper does some experiments to evaluate the performance overhead between the normal mode and the safe mode. The experiments show that the performance overhead of this scheme is reasonable and can effectively reduce the risk of sensitive information leakage.
Conference Paper
Full-text available
In this paper, we demonstrate that Android mal-ware can bypass all automated analysis systems, including AV solutions, mobile sandboxes, and the Google Bouncer. We propose a tool called Sand-Finger for the fingerprinting of Android-based analysis systems. By analyzing the fingerprints of ten unique analysis environments from different vendors, we were able to find characteristics in which all tested environments differ from actual hardware. Depending on the availability of an analysis system, malware can either behave benignly or load malicious code at runtime. We classify this group of malware as Divide-and-Conquer attacks that are efficiently obfuscated by a combination of fingerprinting and dynamic code loading. In this group, we aggregate attacks that work against dynamic as well as static analysis. To demonstrate our approach, we create proof-of-concept malware that surpasses up-to-date malware scanners for Android. We also prove that known malware samples can enter the Google Play Store by modifying them only slightly. Due to Android's lack of an API for malware scanning at runtime, it is impossible for AV solutions to secure Android devices against these attacks.
Article
Full-text available
Antivirus companies, mobile application marketplaces, and the security research community, employ techniques based on dynamic code analysis to detect and analyze mobile malware. In this paper, we present a broad range of anti-analysis techniques that malware can employ to evade dynamic analysis in emulated Android environments. Our detection heuristics span three different categories based on (i) static properties, (ii) dynamic sensor information, and (iii) VM-related intricacies of the Android Emulator. To assess the effectiveness of our techniques, we incorporated them in real malware samples and submitted them to publicly available Android dynamic analysis systems, with alarming results. We found all tools and services to be vulnerable to most of our evasion techniques. Even trivial techniques, such as checking the value of the IMEI, are enough to evade some of the existing dynamic analysis frameworks. We propose possible countermeasures to improve the resistance of current dynamic analysis tools against evasion attempts.
Article
Full-text available
The large amounts of malware, and its diversity, have made it necessary for the security community to use automated dynamic analysis systems. These systems often rely on virtualization or emulation, and have recently started to be available to process mobile malware. Conversely, malware authors seek to detect such systems and evade analysis. In this paper, we present techniques for detecting Android runtime analysis systems. Our techniques are classified into four broad classes showing the ability to detect systems based on differences in behavior, performance, hardware and software components, and those resulting from analysis system design choices. We also evaluate our techniques against current publicly accessible systems, all of which are easily identified and can therefore be hindered by a motivated adversary. Our results show some fundamental limitations in the viability of dynamic mobile malware analysis platforms purely based on virtualization.
Conference Paper
Full-text available
Malicious applications pose a threat to the security of the Android platform. The growing amount and diversity of these applications render conventional defenses largely ineffective and thus Android smartphones often remain un-protected from novel malware. In this paper, we propose DREBIN, a lightweight method for detection of Android malware that enables identifying malicious applications di-rectly on the smartphone. As the limited resources impede monitoring applications at run-time, DREBIN performs a broad static analysis, gathering as many features of an ap-plication as possible. These features are embedded in a joint vector space, such that typical patterns indicative for malware can be automatically identified and used for ex-plaining the decisions of our method. In an evaluation with 123,453 applications and 5,560 malware samples DREBIN outperforms several related approaches and detects 94% of the malware with few false alarms, where the explana-tions provided for each detection reveal relevant properties of the detected malware. On five popular smartphones, the method requires 10 seconds for an analysis on average, ren-dering it suitable for checking downloaded applications di-rectly on the device.
Conference Paper
Full-text available
Smartphones in general and Android in particular are increasingly shifting into the focus of cybercriminals. For understanding the threat to security and privacy it is important for security researchers to analyze malicious software written for these systems. The exploding number of Android malware calls for automation in the analysis. In this paper, we present Mobile-Sandbox, a system designed to automatically analyze Android applications in two novel ways: (1) it combines static and dynamic analysis, i.e., results of static analysis are used to guide dynamic analysis and extend coverage of executed code, and (2) it uses specific techniques to log calls to native (i.e., "non-Java") APIs. We evaluated the system on more than 36,000 applications from Asian third-party mobile markets and found that 24% of all applications actually use native calls in their code.
Thesis
In recent years, the sales volume of smartphones has tremendously increased and the trend changed from old-fashioned mobile phones with limited functionality to powerful smartphones with plenty of features. The Google smartphone platform Android has become the most popular operating system and has overtaken Symbian- and iOS-based smartphones. Smartphones are ubiquitous which is why they increasingly play an important role for evidence in forensic examinations. In this context, the recovery of digital traces is an important factor when examining and clarifying facts of a criminal action. Although there are already some tools and process descriptions existent, a huge demand for methods and tools which are needed for forensic extraction and analysis of data that have been stored on a smartphone can be noticed. This demand is fueled by the rapid growth and increasing diversification of the mobile phone market, too. Due to the high penetration rate as well as the increasing popularity, smartphones are not only used by criminal actions but they become an attractive target for criminals, too. For a better understanding of the related threats to end users it is important to analyze and identify malicious software. The exponentially growing number of Android Malware within the last months demands an automation of the analysis process to cope with the fast growing data volume. Within this thesis the specifications and particularities of Android smartphones are described. Moreover, an extended framework is presented. This framework is able to extract and analyze data, being interesting for criminal investigation, from an SQLite-database as well as system files of Android smartphones ADEL. In addition, it is shown how movement profiles of smartphones users can be generated. With these movement profiles the smartphone’s location can be proven for a certain time span in the past. This step could be particularly important for resolving criminal cases in which a determined time-location-relationship is essential. An example for such a questions is: “Was the suspect close to the robbed villa during the time of burglary?” When resolving crimes, an investigator often comes to the point at which he has to question: “Was the activity on the smartphone made by the owner or has the smartphone been manipulated?”As to provide support to answer this question, the second part of the Thesis focuses on the analysis of mobile malicious code and the analysis system Mobile-Sandbox is presented. This system has been developed to automatically analyze Android applications, in large quantities, too. It combines static and dynamic analysis methods to test the largest possible part of implemented functions of an application. Furthermore, it uses particular techniques to monitor calls from “non-JAVA” API’s. These techniques and the combination of static and dynamic analysis turn the Mobile-Sandbox into a unique and powerful system at present. With the tools and methodical procedure models presented within this Thesis, an investigator as well as security officers in companies receive important assistance for their daily work.
Conference Paper
WebViews allow Android developers to embed a webpage within an application, seamlessly integrating native application code with HTML and JavaScript web content. While this rich interaction simplifies developer support for multiple platforms, it exposes applications to attack. In this paper, we explore two WebView vulnerabilities: excess authorization, where malicious JavaScript can invoke Android application code, and file-based cross-zone scripting, which exposes a device’s file system to an attacker. We build a tool, Bifocals, to detect these vulnerabilities and characterize the prevalence of vulnerable code. We found 6767 applications with WebView-related vulnerabilities (11%11\,\% of applications containing WebViews). Based on our findings, we suggest a modification to WebView security policies that would protect over 60%60\,\% of the vulnerable applications with little burden on developers.
Conference Paper
Android's security framework has been an appealing sub-ject of research in the last few years. Android has been shown to be vulnerable to application-level privilege esca-lation attacks, such as confused deputy attacks, and more recently, attacks by colluding applications. While most of the proposed approaches aim at solving confused deputy at-tacks, there is still no solution that simultaneously addresses collusion attacks. In this paper, we investigate the problem of designing and implementing a practical security framework for Android to protect against confused deputy and collusion attacks. We realize that defeating collusion attacks calls for a rather system-centric solution as opposed to application-dependent policy enforcement. To support our design decisions, we conduct a heuristic analysis of Android's system behavior (with popular apps) to identify attack patterns, classify dif-ferent adversary models, and point out the challenges to be tackled. Then we propose a solution for a system-centric and policy-driven runtime monitoring of communication chan-nels between applications at multiple layers: 1) at the mid-dleware we control IPCs between applications and indirect communication via Android system components. Moreover, inspired by the approach in QUIRE, we establish semantic links between IPCs and enable the reference monitor to ver-ify the call-chain; 2) at the kernel level we realize mandatory access control on the file system (including Unix domain sockets) and local Internet sockets. To allow for runtime, dynamic low-level policy enforcement, we provide a callback channel between the kernel and the middleware. Finally, we evaluate the efficiency and effectiveness of our framework on known confused deputy and collusion attacks, and discuss future directions.