Intelligent Match Merging to Prevent Obfuscation Attacks on Software Plagiarism Detectors

Aus SDQ-Institutsseminar
Version vom 4. Dezember 2023, 10:41 Uhr von Timur Sağlam (Diskussion | Beiträge)
(Unterschied) ← Nächstältere Version | Aktuelle Version (Unterschied) | Nächstjüngere Version → (Unterschied)
Vortragende(r) Nils Niehues
Vortragstyp Masterarbeit
Betreuer(in) Timur Sağlam
Termin Fr 8. Dezember 2023
Vortragsmodus in Präsenz
Kurzfassung The increasing number of computer science students has prompted educators to rely on state-of-the-art source code plagiarism detection tools to deter the submission of plagiarized coding assignments. While these token-based plagiarism detectors are inherently resilient against simple obfuscation attempts, recent research has shown that obfuscation tools empower students to easily modify their submissions, thus evading detection. These tools automatically use dead code insertion and statement reordering to avoid discovery. The emergence of ChatGPT has further raised concerns about its obfuscation capabilities and the need for effective mitigation strategies.

Existing defence mechanisms against obfuscation attempts are often limited by their specificity to certain attacks or dependence on programming languages, requiring tedious and error-prone reimplementation. In response to this challenge, this thesis introduces a novel defence mechanism against automatic obfuscation attacks called match merging. It leverages the fact that obfuscation attacks change the token sequence to split up matches between two submissions so that the plagiarism detector discards the broken matches. Match merging reverts the effects of these attacks by intelligently merging neighboring matches based on a heuristic designed to minimize false positives. Our method’s resilience against classic obfuscation attacks is demonstrated through evaluations on diverse real-world datasets, including undergrad assignments and competitive coding challenges, across six different attack scenarios. Moreover, it significantly improves detection performance against AI-based obfuscation. What sets our method apart is its language- and attack-independence while its minimal runtime overhead makes it seamlessly compatible with other defence mechanisms.