Immunization against harmful fine-tuning attacks
This work addresses the security threat of harmful fine-tuning for LLM safety, providing foundational guidance for defense development in scenarios where defenders lack control over fine-tuning.
The paper tackles the problem of defending large language models against harmful fine-tuning attacks, where safety guards can be removed, by introducing a formal framework called 'Immunization' conditions based on attacker training budgets and establishing guidelines for rigorous defense research.
Large Language Models (LLMs) are often trained with safety guards intended to prevent harmful text generation. However, such safety training can be removed by fine-tuning the LLM on harmful datasets. While this emerging threat (harmful fine-tuning attacks) has been characterized by previous work, there is little understanding of how we should proceed in constructing and validating defenses against these attacks especially in the case where defenders would not have control of the fine-tuning process. We introduce a formal framework based on the training budget of an attacker which we call "Immunization" conditions. Using a formal characterisation of the harmful fine-tuning problem, we provide a thorough description of what a successful defense must comprise of and establish a set of guidelines on how rigorous defense research that gives us confidence should proceed.