JUE: Unveiling and Eliminating Neural Backdoors by Removing Shortcuts

We present a comprehensive framework designed to address the vulnerabilities in deep neural networks arising from backdoor attacks. Our framework is distinguished by four main components: identification of Trojan neurons, Trigger Synthesis, Backdoor Detection, and Backdoor Elimination using Soft Weight Masking for mitigation. Through rigorous experimentation, we have substantiated the efficacy of our approach in both identifying and mitigating backdoor attacks, all while preserving the model's performance on clean data. This makes our method a robust and versatile defense mechanism, adaptable across a variety of neural network architectures and real-world scenarios.