Evading next gen AV using AI
AI agent using reinforcement learning to modify malware and evade ML-based antivirus detection.
Summary
Next-Gen AV Approach
Next-generation antivirus relies on machine learning to generalize to never-before-seen malware. However, machine learning itself is susceptible to attack by other machine learning models.
Adversarial Machine Learning
The presentation demonstrates an AI agent trained through reinforcement learning to modify malware, enabling evasion of ML-based malware detection. Reinforcement learning has achieved superhuman performance in complex domains like Go and Atari games.
Malware Modification Strategy
The AI agent learns through thousands of games against a next-gen AV detector. It discovers functionality-preserving changes that can be applied to Windows PE malware files to bypass detection.
Accessibility
No machine learning background is required to understand the core concepts. Fundamental understanding of malware and Windows PE files is sufficient. The demonstration shows how AI can be used offensively to defeat AI-based defenses in antivirus products.
Related talks
Village Interviews·14 min
The AI Village demonstrates an innovative artificial intelligence setup capable of playing Pokémon autonomously without expensive API token bills.
DragonJAR - Hacking y Seguridad Informática·1 hr 10 min
Chema Alonso
Demonstrates genAI por Hacking & Fun and evaluates key implications for security practitioners.
The Linux Foundation·38 min
Paul Holland · Hewlett Packard Enterprise
Discusses the O[S]PO Organization for the Age of AI and evaluates key implications for security practitioners.