Adversarial machine learning: evaluation of attack models & defense mechanisms

In recent years, there has been a sharp increase in the use of mobile platforms and particularly devices based on the Android operating system. This rapid use of mobile devices has fueled cybercriminals’ interest in developing and sharing malicious software. Machine learning algorithms can be used t...

Full description

Saved in:
Bibliographic Details
Main Authors: Perifanis, Vasileios, Tserpes, Iosif, Περηφάνης, Βασίλειος, Τσερπές, Ιωσήφ
Other Authors: Rizomiliotis, Panagiotis
Language:en_US
Published: 2020
Subjects:
Online Access:http://hdl.handle.net/11610/20142
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:In recent years, there has been a sharp increase in the use of mobile platforms and particularly devices based on the Android operating system. This rapid use of mobile devices has fueled cybercriminals’ interest in developing and sharing malicious software. Machine learning algorithms can be used to detect malware with extremely high performance. However, many of these algorithms, and mainly neural network models, are vulnerable to changes in the input data, known as adversarial examples, capable of leading a model to produce misclassifications. This weakness is one of the major problems that the research community is called upon to solve. This thesis presents the evolution of malicious software for Android-based devices over time and refers to the extraction of an application’s features to detect malicious activity. In addition, ways of detecting malware through machine learning models are being developed, as well as ways in which an attacker can deceive these models. This work focuses on the experimental demonstration of the efficiency of machine learning models for malware detection and the weakness of these models against small changes in the input data. Finally, methods for defending models are being evaluated and special features of adversarial examples are being discussed.