PVDI: Preserving Vital and Disrupting Irrelevant Latent Attentions for Robust Backdoor Defense

Backdoor attacks (hidden triggers that manipulate AI model predictions while keeping normal performance intact) are a serious security threat to deep neural networks (machine learning models with many layers). This paper presents PVDI, a defense method that removes backdoors by selectively preserving important attention patterns (the AI's focus on relevant input features) while disrupting irrelevant ones, successfully reducing attack success rates without hurting the model's normal performance.