Neuropots - Realtime Proactive Defense against Bit-Flip Attacks in Neural Networks

Explore a conference talk on NeuroPots, a novel proactive defense mechanism against bit-flip attacks in deep neural networks. Learn about the innovative "honeypot" approach that embeds crafted vulnerabilities to lure attackers and facilitate efficient detection and model recovery. Discover how this trapdoor-enabled defense framework selects honey neurons, embeds trapdoors, and uses checksum-based detection to protect DNN models. Understand the effectiveness of this method across various DNN models and datasets, offering a promising solution for enhancing the security of neural networks in critical applications like self-driving cars and financial systems.