Neural Cleanse: Identifying and Mitigating Backdoor Attacks in Neural Networks
Explore the detection and mitigation of backdoor attacks in deep neural networks in this IEEE conference talk. Delve into the lack of transparency in DNNs that makes them vulnerable to hidden triggers overriding normal classification. Learn about a robust and generalizable system for identifying backdoors and reconstructing possible triggers. Discover multiple mitigation techniques, including input filters, neuron pruning, and unlearning. Examine the efficacy of these techniques through extensive experiments on various DNNs and against different backdoor injection methods. Gain insights into the security risks posed by backdoor attacks in applications such as biometric authentication systems and self-driving cars. Understand the key intuitions behind detecting backdoors and the design overview of the detection process. Review experiment setups, backdoor detection performance, and a brief summary of mitigation strategies.