Antidote: Post-fine-tuning Safety Alignment for Large Language Models against Harmful Fine-tuning
This addresses a critical safety problem for users of large language models by providing a robust defense against harmful fine-tuning, though it is incremental as it builds on existing pruning techniques.
The paper tackles the vulnerability of safety-aligned large language models to harmful fine-tuning attacks, proposing Antidote, a post-fine-tuning pruning method that reduces harmful scores while maintaining downstream task accuracy, with empirical results demonstrating its effectiveness.
Safety aligned Large Language Models (LLMs) are vulnerable to harmful fine-tuning attacks -- a few harmful data mixed in the fine-tuning dataset can break the LLMs's safety alignment. While several defenses have been proposed, our evaluation shows that existing defenses fail \textit{when some specific training hyper-parameters are chosen} -- a large learning rate or a large number of training epochs in the fine-tuning stage can easily invalidate the defense. To this end, we propose Antidote, a post-fine-tuning stage solution, which remains \textbf{\textit{agnostic to the training hyper-parameters in the fine-tuning stage}}. Antidote relies on the philosophy that by removing the harmful parameters, the harmful model can be recovered from the harmful behaviors, regardless of how those harmful parameters are formed in the fine-tuning stage. With this philosophy, we introduce a one-shot pruning stage after harmful fine-tuning to remove the harmful weights that are responsible for the generation of harmful content. Despite its embarrassing simplicity, empirical results show that Antidote can reduce harmful score while maintaining accuracy on downstream tasks. Code is available at https://github.com/git-disl/Antidote.