Towards a Proactive ML Approach for Detecting Backdoor Poison Samples

Explore a 15-minute conference talk from USENIX Security '23 that presents a proactive machine learning approach for detecting backdoor poison samples in deep learning models. Delve into the researchers' investigation of how to mitigate the threat of backdoor attacks by uncovering and addressing limitations in existing post-hoc defense workflows. Learn about their proposed paradigm shift towards a proactive mindset in poison detection, including a unified framework and practical insights for designing more robust and generalizable detection pipelines. Discover the innovative Confusion Training (CT) technique, which applies an additional poisoning attack to expose backdoor patterns more effectively. Examine the empirical evaluations conducted across multiple datasets and attack types, demonstrating the superiority of this approach over existing baseline defenses.