Agent Deprogramming Finetuning Away Backdoor Triggers for Secure Machine Learning Models/LLMs

Adversarial fine-tuning has shown promise in improving model robustness against test-time perturbations, but its effectiveness in mitigating backdoor attacks in language models remains understudied. This work investigates the use of adversarial fine-tuning techniques and deprogramming, specifically focusing on basic fine-tuning, fine-tuning with paraphrasing, and fine-tuning with retokenization, to mitigate backdoor attacks in language models. We evaluate the effectiveness of these techniques by measuring the accuracy of our models with and without the context of the original query. Our approach extends the idea of supervised fine-tuning (SFT) and incorporates adaptors from the parameter-efficient fine-tuning (PEFT) framework to efficiently fine-tune our models while minimizing computational resources. While supervised fine-tuning alone can alleviate backdoor triggers to some extent, our results demonstrate that combining SFT with paraphrasing and retokenization techniques can achieve even more effective mitigation. In particular, adversarial fine-tuning with paraphrasing significantly reduces the success rate of backdoor attacks in language models.