Reproduce “Evaluating the Robustness of a Production Malware Detection System to Transferable Adversarial Attacks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Milad Nasr, Yanick Fratantonio, Luca Invernizzi, Ange Albertini, Loua Farah, Alex Petit-Bianco, Andreas Terzis, Kurt Thomas and 2 more
This research analyzes the robustness of Google's Magika file-type classifier against transferable adversarial attacks, demonstrating how vulnerabilities in this machine learning component can lead to an end-to-end bypass of Gmail's malware detection system. It introduces a practical AES-based preprocessing defense that substantially reduces attack transferability with minimal impact on accuracy and latency, which has been deployed in a production setting.
Key insights
- Adversarial examples are not merely theoretical; they represent concrete, exploitable vulnerabilities that can compromise end-to-end security pipelines in production systems.
- Simple randomization and adversarial training offer limited or impractical defense against resourceful adversaries due to accuracy trade-offs or susceptibility to ensemble attacks.
- A cryptographic transformation like AES-based preprocessing can effectively disrupt the input-gradient landscape, providing robust protection against transferable attacks with minimal impact on benign accuracy and computational overhead.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.