New Defense Traps AI Backdoors with Decoy Shortcuts

A new defense strategy called Trapping and Removing (TR) has been developed to combat backdoor attacks in deep neural networks, which are particularly concerning when models are trained on third-party data. The core insight behind TR is that backdoor behaviors tend to be captured by simpler, parallel network branches when trained alongside the main network. TR leverages this by introducing a lightweight "shortcut" branch that acts as a honeypot, specifically designed to absorb the malicious backdoor knowledge during training. After the training phase, this shortcut branch can simply be discarded, effectively removing the backdoor without requiring any additional data or complex post-processing. To further enhance the isolation of backdoor knowledge while maintaining the main network's performance on benign tasks, the researchers also devised a knowledge decoupling strategy using entropy-based weight assignment. This encourages poisoned samples to flow through the honeypot while guiding the main network to focus on legitimate learning. An automatic shortcut generation strategy ensures generalization across different model architectures. Extensive experiments across various datasets and architectures confirm TR's effectiveness in mitigating a wide range of backdoor attacks while preserving the model's performance on clean data.