Conference PaperPDF Available

Artificial Intelligence-based Attack and Countermeasure Agents: Who wins? An Invited Paper

Authors:

Figures

Content may be subject to copyright.
Artificial Intelligence-based Attack and
Countermeasure Agents: Who wins?
An Invited Paper
Mee Hong Ling and Kok-Lim Alvin Yau
Department of Computing and Information Systems
Sunway University
Kuala Lumpur, Malaysia
{mhling,koklimy}@sunway.edu.my
Abstract—Reinforcement learning (RL) is a branch of artificial
intelligence that has been well investigated to show system
performance enhancement. Yet, the investigation into the security
aspects of RL is at its infancy. This paper offers an investigation
of the security aspect of RL. In this paper, RL is applied to
cluster size adjustment in clustering for distributed cognitive
radio networks whereby the unlicensed or secondary users (SUs)
explore and use underutilized channels (or white spaces) owned
by the licensed or primary users. Clustering segregates SUs in a
network into logical groups; and each cluster consists of a leader
(or a clusterhead) and member nodes. The investigation covers
the effects of an important RL parameter, specifically the learning
rate α, in a dynamic operating environment. Both clusterhead
and member node leverage on RL: a) the clusterhead uses RL
to countermeasure attacks, and b) the SU uses RL to launch
attacks with various attack probabilities. Simulation results have
shown that a RL model with learning rate α= 1 for clusterhead
provides the highest network scalability when being attacked
with various attack probabilities and different learning rates in
a dynamic operating environment.
Index Terms—Artificial intelligence, machine learning, cogni-
tive radio, clustering, reinforcement learning, security, attacks.
I. INTRODUCTION
The field of artificial intelligence (AI) is in the area of
sustainable development and it has made a significant impact
in the development of digital era, with impressive revolutions
governed by theories and information technology [1], [2]. With
this upward trend of growth in the 21st century, however, the
need to address security issues in AI is relevant in order to
experience a greater chance of success in implementation [3].
Cybersecurity is the main concern for today’s digital world
and the impact of AI in this area is uncertain [4]. While AI
has been seen to automate the detection of threat, assures
complete error-free cyber-security services, and provides the
defenders protection and being resilient even against a series
of attacks [5], [6], the downside of AI is unknown since
most of the features of AI are still uncovered which may
lead to vulnerabilities regarding its usage [7], [8]. While
the defenders leverage on AI to countermeasure attacks, the
attackers likewise may also exploit AI to their benefits. Hence,
the need to investigate AI’s effectiveness in security is vital.
Reinforcement learning (RL) is a branch of AI that is
well investigated to show system performance enhancement
when the agents are operating in a volatile environment. It
has been concluded in [9] that the RL model is an effective
model in clustering to achieve higher network scalability in
distributed cognitive radio networks (DCRNs). DCRNs consist
of member nodes (or secondary users) that launch attacks with
various attack probabilities. The clusterhead, which resides
in an operating region (environment) that is characterized
by the attack probabilities launched by the malicious SUs,
countermeasures by leveraging on a RL model.
While the RL technique has been applied to various sce-
narios to tackle attacks in DCRNs, investigations into the RL
parameter, namely the learning rate, for both the attackers and
clusterhead has not been investigated.
The organization of this article is as follows. Section II
presents the background of our work. It provides discussion
on RL, attackers and clusterhead activities, clustering, and our
contribution. Section III discusses a clustering scenario, and
proposes RL models for a clusterhead and attackers, respec-
tively. Section IV presents the performance metric, simulation
parameters, and simulation results and discussions. Section V
concludes our investigation.
II. BACKGROU ND
The following sub-sections provide an overview of the
various components used in our work, together with our
contribution.
A. Reinforcement Learning
Reinforcement learning (RL) is a branch of machine learn-
ing, and it can be categorized under artificial intelligence due
to its autonomous learning algorithms [10], [11]. RL enables
the agents to observe and learn about the dynamic operating
environment without any guidance from their supervisors, and
make decisions on action selection in order to achieve optimal
or near-optimal system performance. Q-learning [12] is a
single-agent on-line algorithm used in our simulation work.
In Q-learning, each agent observes, learns and acts simul-
taneously. The three main elements in RL are state, action
and reward. The state is the agent’s decision making factor,
which is observed or derived from the operating environment.
The reward is the agent’s optimal or near optimal performance
ALGORITHM 1: RL Algorithm
1: procedure
2: Observe current state si
t
3: if exploration then
4: Select a random action ai
t
5: else
6: Select an optimal action ai,
tusing Eq. (1)
7: end if
8: Receive delayed reward ri
t+1(si
t+1, ai
t+1)
9: Update Q-value Qi
t+1(si
t, ai
t)using Eq. (2)
10: end procedure
metrics, such as higher network scalability. The action is the
agent’s selected action taken in order to maximize its reward.
In a RL approach, each agent or decision maker observes
and learns from the operating environment independently. In
our work, we embed RL in the clusterhead and in the attackers.
The RL algorithm maximizes the local Q-value as shown
respectively in Algorithm 1. An agent iobserves state si
tS
which is the local decision making factors from the dynamic
operating environment and selects an action ai
tAat decision
epoch t. The agent chooses either to exploit or explore its
action. It exploits when it takes a greedy action with the
maximum Q-value which represents the appropriateness of a
state-action pair, as follows:
ai,
t= argmax
aA
Qi
t(si
t, a)(1)
An agent explores when it selects a random action in order
to update the Q-values of non-greedy actions in an attempt to
discover better actions. Upon action selection, an agent ire-
ceives a delayed reward ri
t+1(si
t+1, ai
t+1)which represents the
positive or negative effect from its local operating environment
at the next decision epoch t+ 1 [11]. As the decision epoch
continues t= 1,2, . . . , an agent iexplores all state-action
pairs (si
t, ai
t)and updates their respective Q-values Qi
t(si
t, ai
t)
using Q-function as follows:
Qi
t+1(si
t, ai
t)(1 α)Qi
t(si
t, ai
t)
+α[ri
t+1(si
t+1) + γmax
aAQi
t(si
t+1, a)] (2)
Note that, there are two tunable parameters in Q-function,
namely the learning rate αand the discount factor γ, and both
parameters range from 0 to 1. The learning rate αdescribes the
relationship between the old and the newly learnt knowledge.
It determines the rate in which the new estimates or knowledge
are to be incorporated into the old ones. For instance, learning
rate α= 0 causes an agent inot to learn from the operating
environment resulting in using the old knowledge in decision
making, while learning rate α= 1 causes an agent ito use the
newly learnt knowledge. The discount factor γdetermines how
much effect the future rewards have on the optimal decisions.
B. Attacker and Clusterhead
In our work, we focus on two main entities, namely the
attackers (malicious SUs) and the legitimate clusterhead.
Each attacker has its capability to launch attacks with certain
attack probability 0Pi
t0.9. Since it is the nature of
the attackers to be covert, Pi
t= 1 is not considered. In the
attack scenario, the attackers leverage on a RL model to learn
the clusterhead’s perception about their behavior, which is
represented by a state. Using RL, the attackers can compute
the clusterhead’s Q-values and launch their subsequent attacks,
which is represented by action, in order not to be perceived
as malicious. The positive effect of the attackers’ actions (i.e.,
delayed reward) is a decrease in the cluster size.
As a countermeasure, the clusterhead adopts a RL model
to learn its member nodes’ behavior, which is represented
by Q-values. These values is used to keep track of the
nodes’ behavior. Using the Q-values, the clusterhead selects
the legitimate nodes represented by action. For instance, when
the clusterhead observes that its selected member node (i.e.,
action) utilizes the granted resources, its Q-value increases.
The positive effect of the clusterhead’s action (i.e., delayed
reward) is an increase in the cluster size.
C. Clustering
As some of the PUs’ transmission range may not cover the
entire cognitive radio networks topology, with some SUs not
located within the PUs’ range, a desirable approach to maxi-
mize the unutilized resources is through clustering. Clustering
allows collaboration amongst the SUs [13], and it organizes
SUs into logical groups (clusters) based on clustering param-
eters such as the number of common available channels in
a cluster. A node in a cluster is either a member node or a
clusterhead. A clusterhead acts as a local node that provides
vital cognitive radio operations such as routing while a mem-
ber node is attached to the clusterhead. Several advantages of
using clustering are to increase routing efficiency by having
minimum nodes in the backbone network and consequently
to reduce update on routing information cost [14]. Hence,
leveraging on clustering increases white spaces utilization and
thereby provides greater network scalability [15].
D. Our Contribution
This paper specifically investigates the effects of RL’s
learning rate αof both the attackers and the clusterhead on the
network scalability (or cluster size ratio) in a volatile operating
environment.
III. PROP OS ED R L MODELS
This section describes the clustering scenario and the pro-
posed RL models for both the attackers and the clusterhead.
Fig. 1 shows an abstract view of RL for both the attacker and
the clusterhead. The explanation of this diagram can be found
in the following sub-sections. The general notation used can
be found in Table I.
Fig. 1: An abstract view of RL attacker and clusterhead
TABLE I: General notations
Notation Description
tTA decision epoch (also called a time window) with
T={1,2, . . . }.
tsTsA time slot with Ts={1,2,...,|Ts|}.
mMA PU with M={1,2,...,|M|}. PU moccupies
licensed channel m.
i∈ I A SU with I={1,2,...,|I|}.
Λm,tsAvailability of channel mat decision epoch t.
Λm,ts= 1 means PU is idle or OFF at time slot
ts, and Λm,ts= 0 means PU is busy or ON at
time slot ts.
Λi
m,tsEquivalent of Λm,tsfor SU i.
Pm,tsTransition probability matrix of PU activity in
channel mat time slot ts.
pdb
m,tspdb
m,ts=Prob{Λm,ts= 0|Λm,ts1= 1}is the
probability of channel mswitches from state OFF
at time slot ts1to ON at time slot t.
pbd
m,tspbd
m,ts=Prob{Λm,ts= 1|Λm,ts1= 0}is the
probability of channel mswitches from state ON
at time slot ts1to OFF at time slot t.
Pi
tAttack probability launched by SU iat decision
epoch t.
PcAttack probability of a cluster c.
A. Clustering scenario
At the start of each decision epoch t, a clusterhead is given
certain amount of unutilized white spaces, and it attempts
to distribute this resource to its one-hop neighboring nodes.
The aim of the clusterhead is to maximize its cluster size
based on the given resources. RL algorithm is embedded in
the clusterhead so that at the end of each decision epoch t
(i.e., after data packet transmission), the clusterhead learns its
member nodes’ behavior through Q-value updates. Higher Q-
values indicate more dependable or legitimate nodes —this
facilitates the selection of nodes. In our work, we consider
the clusterhead to be legitimate at all times and its cluster size
at each decision epoch tis influenced by the given amount
of white spaces, and that the clusterhead does not waste any
resources. In addition, there are sufficient number of SUs (with
certain attack probability Pi
t) that wish to become members in
TABLE II: RL model embedded in attacker i
State si
tS= (1,2,... 4) represents the performance indicator
for attacker iat decision epoch t. States si
t= 1 and
si
t= 4 indicate that the attacker has the worst and the best
performance, respectively
Action ai
tA= (Pi
t)represents the attack probability Pi
t, where
0.1Pi
t0.9
Reward ri
trepresents the unused white spaces
the cluster. Throughout decision epoch T, a cluster is assumed
to have a maximum average attack probability Pc, which is
derived from the initial mean of the attack probability Pi
0of
all its attackers as shown in (3). Unless otherwise stated, the
attack probability discussed in this paper refers to the cluster’s
attack probability, namely Pc.
Pc=
|I|
P
i=1
Pi
0
|I| (3)
B. Proposed RL model for attacker
The RL model for the attacker is shown in Table II.
Each malicious SU ilaunches attacks independently with the
objective to decrease the cluster size. This is carried out by
wasting the requested/granted resources (white spaces) —by
not using the resources for data packet transmission and/or
help the clusterhead to distribute the resources to its down-
stream nodes. Specifically, the malicious SU maximizes the
unused resources and aims to be covert at all times, and yet to
have continued supply of resources from the clusterhead. This
attack, which is also known as an intelligent attack, enables
the malicious SU ito learn from the operating environment by
observing its current performance indicator Di
t[0,1], which
is calculated based how well the clusterhead perceives the
malicious SU. Higher perceived performance indicator value
indicates greater perceived trustworthiness of the malicious
SU. The performance indicator Di
tis uniformly partitioned
into four sub-ranges as seen in (4), and so the state of each
malicious SU can be represented as follows:
si
t=
1 (poor) 0 Di
t<0.25,
2 (average) 0.25 Di
t<0.5,
3 (good) 0.5Di
t<0.75,
4 (best) 0.75 Di
t1.
(4)
The attacker’s action ai
tis denoted by the attack probability,
0Pi
t0.9and the delayed reward ri
tof a malicious SU i
refers to the wasted white spaces.
C. Proposed RL model for clusterhead
The RL model for the clusterhead is shown in Table III.
Note that in our scenario, the clusterhead jdoes not have a
state as its actions do not have any effect on its subsequent
states —the resources given to the clusterhead at each decision
epoch tis affected by the PUs’ activities and not by the
TABLE III: RL model embedded in a clusterhead j
Action aj
tA={aj
t
max Qj
t(aj
t)}where j={1,2,...,n},
represents set of one-hop member nodes in a cluster node
with the highest Q-value is chosen, followed by a node with
the second highest Q-value, and so on.
Reward rj
trepresents the total amount of used white spaces by the
chosen member nodes at decision epoch t
clusterhead’s action. Hence, in a stateless RL model, the
clusterhead learns to optimize its reward solely based on the
best action taken. Denote aj
tA={aj
t
max Qj
t(aj
t)}where
j={1,2, . . . , n}, is a set of one-hop member nodes with
the highest Q values in a cluster. Upon action selection, the
delayed reward is rj
t, which is the used white space after the
each chosen node has added new nodes and/or transmitted its
data packets and/or relay packets for its downstream nodes.
The clusterhead then updates its member nodes’ Q-values
accordingly.
IV. PERFORMANCE EVALUATION,RESULTS A ND
DISCUSSIONS
A. Performance metric
At the end of each decision epoch t, the clusterhead’s
performance is measured based on cluster size ratio that
compares the cluster sizes before and after the utilization
of resources as shown in (5). As the clusterhead aims to
achieve greater network scalability, higher cluster size ratio
Rc,t is desirable. Eq. 5 shows that Zc,ts=|Ts|is the cluster
size measured at the end of a decision epoch tafter an attack
has taken place, and Zc,ts=1 is the cluster size measured at the
start of a decision epoch t when the clusterhead has distributed
the resources to its member nodes.
Rc,t =Zc,ts=|Ts|
Zc,ts=1
(5)
B. Simulation settings
The simulation is done using MATLAB. For each round
of simulation, we generate a new network topology using the
MATLAB random generator.
C. Simulation parameters
The simulation parameters are shown in Table IV. As the
primary investigation of this paper is on the learning rate, two
primary users |M|= 2 suffice in the DCRN. To re-iterate, we
provide some uncertainty in the attack, and hence the attack
probability, Pi
t= 1 is not considered in this paper.
D. Simulation Results and Discussions
Fig. 2 and Fig. 3 present the performance of RL attackers
and RL clusterhead, with both agents adjusting their learning
rate to attack and countermeasure, respectively. In Fig. 2,
the learning rates for attackers (α= 1) and clusterhead
α= (0.5,1) are plotted against the attack probability (x-axis)
and cluster size ratio (y-axis), while in Fig. 3, the learning rates
TABLE IV: Simulation parameters
Notation Value
T{1,2,...,10000}
|M|2
P0.5 0.5
0.5 0.5
|I| {1,2,...,50}
Pi
t{0.1,0.2,...,0.9}
Pc{0.1,0.2,...,0.9}
α0.5,1
for attackers α= (0.5,1) and clusterhead (α= 1) are also
plotted against the attack probability (x-axis) and cluster size
ratio (y-axis). Both graphs show that the varying learning rate
values for attackers and clusterhead have a similar trend —the
cluster size ratio decreases as the attack probability increases.
There are two observations made from these graphs as follows:
1) when both attackers and clusterhead learn at the fastest
speed, namely α= 1, as seen in Fig. 2, there is an in-
crease of about 10% in performance for the clusterhead
as compared to the clusterhead that adopts α= 0.5. This
shows that in any attack and countermeasure scenarios,
the need to learn at the fastest rate is crucial.
2) when the attackers learn at a slower rate α= 0.5as
seen Fig. 3, they can also achieve similar performance
as when they adopt learning rate α= 1. This is due to
the covert nature of attackers —they launch attacks with
varying probabilities.
Fig. 2: Analysis of cluster size ratio with varying learning rate
αfor clusterhead
V. C ON CL US IO N AN D FU RTH ER W OR K
Our simulation work was based on both RL (intelligent)
attacks and countermeasures. The results have proven that in
any volatile environment, both the attackers and clusterhead
need to learn fast in order to win. Due to the covert nature
Fig. 3: Analysis of cluster size ratio with varying learning rate
αfor attackers
of the attackers, further work can be carried out to enhance
the cluster size adjustment scheme for the clusterhead. For
instance, the clusterhead may deploy a rule-based RL that
only grants resources to its member nodes of certain Q-value
while maintaining its maximum learning rate. In addition,
deep learning, which is a new area of machine learning, has
proven to yield better results [16], [17], can also be explored
to increase the performance of the clusterhead.
REFERENCES
[1] K.F. Lee, “AI Superpowers: China, Silicon Valley, and the New World
Order,” Boston, Mass: Houghton Mifflin. ISBN 9781328546395. OCLC
1035622189.
[2] J. Liu, X. Kong, F. Xia, X. Bai, L. Wang, Q. Qing and I. Lee, “Artificial
Intelligence in the 21st Century,IEEE Access, vol. 6, pp. 34403-34421,
2018.
[3] M. Hutson, “Hackers easily fool artificial intelligences,” Science, vol.
361, no. 6399, pp. 215-215, 2018.
[4] R.V. Yampolskiy, “Predicting future AI failures from historic examples,”
Foresight, vol. 21 no. 1, pp.138-152, 2019.
[5] R.V. Yampolskiy, M.S. Spellchecker, “Artificial Intelligence Safety and
Cybersecurity: a Timeline of AI Failures,CoRR, vol. abs/1610.07997,
2016.
[6] S. Greengard, “Cybersecurity Gets Smart,” Commun. ACM, vol. 59, no.
5, pp 29-31, 2016.
[7] A.S. Wilner, “Cybersecurity and its discontents: Artificial intelligence,
the Internet of Things, and digital misinformation,” International Jour-
nal, SAGE Publications vol. 73, no. 2 pp 308–316 2018.
[8] S. Berberich, “Artificial intelligence has a role in cybersecurity but isn’t
the magic bullet,” Rochester Business Journal, SAGE Publications vol.
34, no. 31, 2018.
[9] M.H. Ling and K.L.A. Yau, “Can Reinforcement learning address
security issues? an investigation into a clustering scheme in distributed
cognitive radio networks, The 33rd International Conference on Infor-
mation Networking (ICOIN-2019), Jan 2019.
[10] M.L. Littman, “Reinforcement learning improves behaviour from evalu-
ative feedback,Nature, Nature Publishing Group, vol. 521, no. 13, pp.
445-451, 2015.
[11] R.S. Sutton, and A.G. Barto, Reinforcement Learning: An Introduction.
Cambridge, MA, USA: MIT Press; May 1998.
[12] C. J. C. H. Watkins,“Learning from Delayed Rewards,Cambridge
University, 1989.
[13] M. Zheng, S. Chen, W. Liang and M. Song, “NSAC: A Novel Clustering
Protocol in Cognitive Radio Sensor Networks for Internet of Things”,
IEEE Internet of Things Journal, doi=10.1109/JIOT.2019.2898166,
ISSN=2327-4662
[14] G. Sharma and R. Sharma, “Cluster-based distributed cooperative spec-
trum sensing over Nakagami fading using diversity reception,IET
Networks, ISSN = 2047-4954, January 2019.
[15] H. Idoudi, O. Mabrouk, P. Minet and L.A. Saidane, “Cluster-based
scheduling for cognitive radio sensor networks,Journal of Ambient
Intelligence and Humanized Computing, vol. 10, no. 2, pp 477 -489,
2019.
[16] G.E. Hinton, S. Osindero, and Y.W. Teh, “ A fast learning algorithm
for deep belief nets,” Neural Comput., vol. 18, no. 7, pp. 1527-1554,
2006.
[17] Y. LeCun, Y. Bengio, and G.E. Hinton, “ Deep learning,” Nature., vol.
521, pp. 436-444, 2015.
ResearchGate has not been able to resolve any citations for this publication.
Article
Full-text available
Cooperative spectrum sensing (CSS) can solve the problem of hidden terminal in cognitive radio (CR). CSS can be employed in two ways distributed and centralised. In wireless networking, clustering can provide network scalability, better resource allocation, and energy efficiency. Here, cluster‐based distributed CSS is investigated over Nakagami channel. In the proposed scheme, four fusion policies OR–OR, OR–AND, AND–OR, and AND–AND are investigated. An analytical framework has been presented to evaluate different parameters related to spectrum sensing, i.e. detection probability, false alarm probability, and missed detection probability for Nakagami fading channel. On the basis of the developed framework, the performance of cluster‐based distributed CSS has been compared with the centralised CSS. Results show that OR–OR fusion rule of the cluster‐based distributed CSS outperforms other fusion rules and centralised CSS. To achieve 80% probability of detection, SNR required is 10 dB for other fusion schemes. Further performance improvement is achieved by using square law selection and maximum ratio combining diversity schemes.
Article
Full-text available
Purpose The purpose of this paper is to explain to readers how intelligent systems can fail and how artificial intelligence (AI) safety is different from cybersecurity. The goal of cybersecurity is to reduce the number of successful attacks on the system; the goal of AI Safety is to make sure zero attacks succeed in bypassing the safety mechanisms. Unfortunately, such a level of performance is unachievable. Every security system will eventually fail; there is no such thing as a 100 per cent secure system. Design/methodology/approach AI Safety can be improved based on ideas developed by cybersecurity experts. For narrow AI Safety, failures are at the same, moderate level of criticality as in cybersecurity; however, for general AI, failures have a fundamentally different impact. A single failure of a superintelligent system may cause a catastrophic event without a chance for recovery. Findings In this paper, the authors present and analyze reported failures of artificially intelligent systems and extrapolate our analysis to future AIs. The authors suggest that both the frequency and the seriousness of future AI failures will steadily increase. Originality/value This is a first attempt to assemble a public data set of AI failures and is extremely valuable to AI Safety researchers.
Article
Full-text available
Last week, at the International Conference on Machine Learning (ICML) in Stockholm, a group of researchers described a turtle they had 3D printed. Most people would say it looks just like a turtle, but an artificial intelligence (AI) algorithm that can normally recognize turtles saw it differently. Most of the time, it thought the turtle was a rifle. Similarly, it saw a 3D-printed baseball as an espresso. These are examples of "adversarial attacks"—subtly altered images, objects, or sounds that fool AIs without setting off human alarm bells. Impressive advances in AI—particularly machine learning algorithms that can recognize sounds or objects after digesting vast training data sets—have spurred the growth of living room voice assistants and autonomous cars. But these AIs are surprisingly vulnerable to being spoofed. At the ICML meeting, adversarial attacks were a hot subject, with researchers reporting novel ways to trick AIs as well as new ways to defend them. Somewhat ominously, one of the conference's two best paper awards went to a study suggesting protected AIs aren't as secure as their developers might think.
Article
Full-text available
The field of Artificial Intelligence (AI) has shown an upward trend of growth in the 21st Century (from 2000 to 2015). The evolution in AI has advanced the development of human society in our own time, with dramatic revolutions shaped by both theories and techniques. However, the multidisciplinary and fast-growing features make AI a field difficult to be well understood. In this paper, we study the evolution of AI at the beginning of the 21st Century using publication metadata extracted from 9 top-tier journals and 12 top-tier conferences of this discipline. We find that the area is in sustainable development and its impact continues to grow. From the perspective of reference behavior, the decrease in self-references indicates that AI is becoming more and more open-minded. The influential papers/researchers/institutions we identified outline landmarks in the development of this field. Last but not least, we explore the inner structure in terms of topics evolution over time. We have quantified the temporal trends at the topic level and discovered the inner connection among these topics. These findings provide deep insights into the current scientific innovations, as well as shedding light on funding policies.
Article
Full-text available
In this paper, we define a cluster based scheduling algorithm for Cognitive Radio Sensor Networks. To avoid inter-clusters collision, we assign fixed channels only to nodes having one-hop neighbors out of their clusters. We denote these nodes as specific nodes. Previous studies assign distinct channels to whole neighbor clusters to avoid inter-clusters collision. Our objective is to optimize the spatial reuse and to increase the network throughput while saving sensors energy. We start by assigning channels only to the specific nodes. Once the problem of inter-clusters collision is solved, each cluster head (CH) schedules the transmissions in its cluster independently. For the cluster members that are specific nodes, the CH assigns only time slots because the channel assignment is already done. For other cluster members (CMs) (not specific nodes), the CH assigns the pair (channel, slot). Two solutions are proposed in this paper to schedule the CMs: the Frame Intra Cluster Multichannel Scheduling algorithm denoted Frame-ICMS and the Slot Intra Cluster Multichannel Scheduling algorithm denoted Slot-ICMS. We evaluate the performance of these algorithms in case of accurate PUs activity detection and in case of bad PUs activity estimation. We prove that our proposals outperform an existing one especially in terms of energy saving.
Article
Full-text available
In this work, we present and analyze reported failures of artificially intelligent systems and extrapolate our analysis to future AIs. We suggest that both the frequency and the seriousness of future AI failures will steadily increase. AI Safety can be improved based on ideas developed by cybersecurity experts. For narrow AIs safety failures are at the same, moderate, level of criticality as in cybersecurity, however for general AI, failures have a fundamentally different impact. A single failure of a superintelligent system may cause a catastrophic event without a chance for recovery. The goal of cybersecurity is to reduce the number of successful attacks on the system; the goal of AI Safety is to make sure zero attacks succeed in bypassing the safety mechanisms. Unfortunately, such a level of performance is unachievable. Every security system will eventually fail; there is no such thing as a 100% secure system.
Article
Clustering is an effective method to manage communications in cognitive radio sensor networks (CRSNs). This letter proposes a network stability-aware clustering (NSAC) protocol for CRSNs. Spectrum dynamics and energy consumption are for the first time simultaneously integrated into the protocol design of NSAC. Extensive simulations show that the proposed NSAC protocol obviously outperforms existing methods in the aspects of network stability and energy consumption.
Article
The future of cybersecurity is in flux. Artificial intelligence challenges existing notions of security, human rights, and governance. Digital misinformation campaigns leverage fabrications and mistruths for political and geostrategic gain. And the Internet of Things—a digital landscape in which billions of wireless objects from smart fridges to smart cars are tethered together—provides new means to distribute and conduct cyberattacks. As technological developments alter the way we think about cybersecurity, they will likewise broaden the way governments and societies will have to learn to respond. This policy brief discusses the emerging landscape of cybersecurity in Canada and abroad, with the intent of informing public debate and discourse on emerging cyber challenges and opportunities.
Article
Researchers aim to apply artificial intelligence and machine-learning methods to take cybersecurity to a new, higher, and better level.