Deep Backdoors in Deep Reinforcement Learning Agents - Understanding Security Threats and Countermeasures

Explore critical security vulnerabilities in Deep Reinforcement Learning (DRL) systems through this 31-minute Black Hat conference presentation. Learn how the complex training requirements and opacity of AI models create opportunities for supply chain attacks, particularly in outsourced training scenarios and pre-trained models from online repositories. Examine sophisticated backdoor embedding techniques demonstrated through practical examples, progressing from basic scenarios to critical applications in autonomous driving, healthcare, and cybersecurity. Understand how attackers can covertly compromise advanced DRL systems while evading detection, and discover real-time detection and neutralization strategies to protect AI-driven technologies. Gain insights from leading experts including Vasilios Mavroudis from The Alan Turing Institute, Jamie Gawith from the University of Bath, Sañyam Vyas from Cardiff University, and Chris Hicks from the AICD Research Centre, as they reveal the hidden dangers and defensive measures in the rapidly evolving landscape of DRL security.