Adversarial Machine Learning for the Protection of Legitimate Software
Date of Award
Jeffrey Todd McDonald, Ph.D.
Obfuscation is the transforming a given program into one that is syntactically different but semantically equivalent. This new obfuscated program now has its code and/or data changed so that they are hidden and difficult for attackers to understand. Obfuscation is an important security tool and used to defend against reverse engineering. When applied to a program, different transformations can be observed to exhibit differing degrees of complexity and changes to the program. Recent work has shown, by studying these side effects, one can associate patterns with different transformations. By taking this into account and attempting to profile these unique side effects, it is possible to create a classifier using machine learning which can analyze transformed software and identifies what transformation was used to put it in its current state. This has the effect of weakening the security of obfuscating transformations used to protect legitimate software. In this research, we explore options to increase the robustness of obfuscation against attackers who utilize machine learning, particular those who use it to identify the type of obfuscation being employed. To accomplish this, we segment our research into three stages. For the first stage, we implement a suite of classifiers that are used to xiv identify the obfuscation used in samples. These establish a baseline for determining the effectiveness of our proposed defenses and make use of three varied feature sets. For the second stage, we explore methods to evade detection by the classifiers. To accomplish this, attacks setup using the principles of adversarial machine learning are carried out as evasion attacks. These attacks take an obfuscated program and make subtle changes to various aspects that will cause it to be mislabeled by the classifiers. The changes made to the programs affect features looked at by our classifiers, focusing mainly on the number and distribution of opcodes within the program. A constraint of these changes is that the program remains semantically unchanged. In addition, we explore a means of algorithmic dead code insertion in to achieve comparable results against a broader range of classifiers. In the third stage, we combine our attack strategies and evaluate the effect of our changes on the strength of obfuscating transformations. We also propose a framework to implement and automate these and other measures. We the following contributions: 1. An evaluation of the effectiveness of supervised learning models at labeling obfuscated transformations. We create these models using three unique feature sets: Code Images, Opcode N-grams, and Gadgets. 2. Demonstration of two approaches to algorithmic dummy code insertion designed to improve the stealth of obfuscating transformations against machine learning: Adversarial Obfuscation and Opcode Expansion 3. A unified version of our two defenses capable of achieving effectiveness against a broad range of classifiers, while also demonstrating its impact on obfuscation metrics.
Parker, Colby, "Adversarial Machine Learning for the Protection of Legitimate Software" (2022). Theses and Dissertations. 35.