Adversarial machine learning in digital forensics

Description

Digital forensics becomes very important when issues about authors of documents arise, such as their identity and characteristics (age, gender) and ability to associate them with unknown documents.

Machine learning approaches to source code authorship identification attempt to identify the most likely author of a piece of code by analyzing various characteristics from source code. There are many situations in which police or security agencies are concerned about the ownership of software, for example, to identify who wrote a malicious piece of code.

However, machine learning models are often susceptible to adversarial deception of their input at test time, which is leading to a poorer performance. Recent studies in adversarial machine learning showed that adversarial examples can easily attack image classification, speech recognition, and reinforcement learning.

In this session we will investigate the feasibility of deception in source code attribution techniques in real world environment, which contains adversaries and dishonest users.

In this session, we will show that even a sensible transformation of author’s coding style successfully decrease the performance of source code authorship attribution systems. An important part of this session’s content will include practical attacks on current attribution systems: author imitation and author hiding. The first attack can be applied on user identity in open-source projects. The attack transforms attacker’s source code to a version that mimics the victim’s coding style while retaining functionality of original code. This is particularly concerning for open-source contributors who are unaware of the fact that by contributing to open-source projects they reveal identifiable information that can be used to their disadvantage. For example, one can easily see that by imitating someone’s coding style it is possible to implicate any software developer in wrongdoing. To resist this attack we discuss multiple approaches of hiding a coding style of software author before contribute to open-source.