Compile-Time Backdoor Neutralization via LLM-Guided Localized Edits
Deep learning compilers such as TVM have become essential in deploying neural networks with high efficiency across diverse hardware. However, recent studies reveal a critical vulnerability: adversaries can embed backdoors during training and rely on compilation to re-materialize the hidden behavior in the final executa...