-
[show abstract]
[hide abstract]
ABSTRACT: Machine Learning algorithms have provided core functionality to many application domains - such as bioinformatics, computational linguistics, etc. However, it is difficult to detect faults in such applications because often there is no "test oracle" to verify the correctness of the computed outputs. To help address the software quality, in this paper we present a technique for testing the implementations of machine learning classification algorithms which support such applications. Our approach is based on the technique "metamorphic testing", which has been shown to be effective to alleviate the oracle problem. Also presented include a case study on a real-world machine learning application framework, and a discussion of how programmers implementing machine learning algorithms can avoid the common pitfalls discovered in our study. We also conduct mutation analysis and cross-validation, which reveal that our method has high effectiveness in killing mutants, and that observing expected cross-validation result alone is not sufficiently effective to detect faults in a supervised classification program. The effectiveness of metamorphic testing is further confirmed by the detection of real faults in a popular open-source classification program.
Journal of Systems and Software 04/2011; 84(4):544-558. · 0.84 Impact Factor
-
[show abstract]
[hide abstract]
ABSTRACT: Many software security vulnerabilities only reveal themselves under certain conditions, i.e., particular configurations of the software together with its particular runtime environment. One approach to detecting these vulnerabilities is fuzz testing, which feeds a range of randomly modified inputs to a software application while monitoring it for failures. However, typical fuzz testing makes no guarantees regarding the syntactic and semantic validity of the input, or of how much of the input space will be explored. To address these problems, in this paper we present a new testing methodology called configuration fuzzing. Configuration fuzzing is a technique whereby the configuration of the running application is randomly modified at certain execution points, in order to check for vulnerabilities that only arise in certain conditions. As the application runs in the deployment environment, this testing technique continuously fuzzes the configuration and checks "security invariants" that, if violated, indicate a vulnerability; however, the fuzzing is performed in a duplicated copy of the original process, so that it does not affect the state of the running application. In addition to discussing the approach and describing a prototype framework for implementation, we also present the results of a case study to demonstrate the approach's efficiency.
Proceedings / The ... International Conference on Availability, Reliability and Security. International Conference on Availability, Reliability and Security. 02/2010;
-
ARES 2010, Fifth International Conference on Availability, Reliability and Security, 15-18 February 2010, Krakow, Poland; 01/2010
-
[show abstract]
[hide abstract]
ABSTRACT: For large, complex software systems, it is typically impossible in terms of time and cost to reliably test the application in all possible execution states and configurations before releasing it into production. One proposed way of addressing this problem has been to continue testing and analysis of the application in the field, after it has been deployed. A practical limitation of many such automated approaches is the potentially high performance overhead incurred by the necessary instrumentation. However, it may be possible to reduce this overhead by selecting test cases and performing analysis only in previously-unseen application states, thus reducing the number of redundant tests and analyses that are run. Solutions for fault detection, model checking, security testing, and fault localization in deployed software may all benefit from a technique that ignores application states that have already been tested or explored.In this paper, we present a solution that ensures that deployment environment tests are only executed in states that the application has not previously encountered. In addition to discussing our implementation, we present the results of an empirical study that demonstrates its effectiveness, and explain how the new approach can be generalized to assist other automated testing and analysis techniques intended for the deployment environment.
Proceedings of the ... International Workshop on Automation of Software Test. International Workshop on Automation of Software Test. 01/2010; 2010:16-23.
-
[show abstract]
[hide abstract]
ABSTRACT: Many software security vulnerabilities only reveal themselves under certain conditions, i.e., particular configurations and inputs together with a certain runtime environment. One approach to detecting these vulnerabilities is fuzz testing. However, typical fuzz testing makes no guarantees regarding the syntactic and semantic validity of the input, or of how much of the input space will be explored. To address these problems, we present a new testing methodology called Configuration Fuzzing. Configuration Fuzzing is a technique whereby the configuration of the running application is mutated at certain execution points, in order to check for vulnerabilities that only arise in certain conditions. As the application runs in the deployment environment, this testing technique continuously fuzzes the configuration and checks "security invariants" that, if violated, indicate a vulnerability. We discuss the approach and introduce a prototype framework called ConFu (CONfiguration FUzzing testing framework) for implementation. We also present the results of case studies that demonstrate the approach's feasibility and evaluate its performance.
International journal of secure software engineering (Print). 01/2010; 1(3):41-55.
-
[show abstract]
[hide abstract]
ABSTRACT: Many applications in the field of scientific computing - such as computational biology, computational linguistics, and others - depend on Machine Learning algorithms to provide important core functionality to support solutions in the particular problem domains. However, it is difficult to test such applications because often there is no "test oracle" to indicate what the correct output should be for arbitrary input. To help address the quality of such software, in this paper we present a technique for testing the implementations of supervised machine learning classification algorithms on which such scientific computing software depends. Our technique is based on an approach called "metamorphic testing", which has been shown to be effective in such cases. More importantly, we demonstrate that our technique not only serves the purpose of verification, but also can be applied in validation. In addition to presenting our technique, we describe a case study we performed on a real-world machine learning application framework, and discuss how programmers implementing machine learning algorithms can avoid the common pitfalls discovered in our study. We also discuss how our findings can be of use to other areas outside scientific computing, as well.
Proceedings / International Conference on Quality Software. International Conference on Quality Software. 01/2010; 2009(2009):135-144.
-
[show abstract]
[hide abstract]
ABSTRACT: Health care professionals rely on software to simulate anatomical and physiological elements of the human body for purposes of training, prototyping, and decision making. Software can also be used to simulate medical processes and protocols to measure cost effectiveness and resource utilization. Whereas much of the software engineering research into simulation software focuses on validation (determining that the simulation accurately models real-world activity), to date there has been little investigation into the testing of simulation software itself, that is, the ability to effectively search for errors in the implementation. This is particularly challenging because often there is no test oracle to indicate whether the results of the simulation are correct. In this paper, we present an approach to systematically testing simulation software in the absence of test oracles, and evaluate the effectiveness of the technique.
Technical Reports (CIS).
-
[show abstract]
[hide abstract]
ABSTRACT: Some machine learning applications are intended to learn properties of data sets where the correct answers are not already known to human users. It is challenging to test and debug such ML software, because there is no reliable test oracle. We describe a framework and collection of tools aimed to assist with this problem. We present our findings from using the testing framework with three implementations of an ML ranking algorithm (all of which had bugs).
-
[show abstract]
[hide abstract]
ABSTRACT: We are concerned with the problem of detecting bugs in machine learning applications. In the absence of sufficient real-world data, creating suitably large data sets for testing can be a difficult task. To address this problem, we have developed an approach to creating data sets called "parameterized random data generation". Our data generation framework allows us to isolate or combine different equivalence classes as desired, and then randomly generate large data sets using the properties of those equivalence classes as parameters. This allows us to take advantage of randomness but still have control over test case selection at the system testing level. We present our findings from using the approach to test two different machine learning ranking applications.
