BEEAR: Embedding-based Adversarial Removal of Safety Backdoors in Instruction-tuned Language Models
This addresses a critical security challenge for AI safety in LLMs, though it is an incremental step towards practical defenses.
The paper tackles the problem of safety backdoor attacks in large language models, which stealthily trigger unsafe behaviors, by proposing BEEAR, a mitigation method that reduces attack success rates from >95% to <1% for RLHF time backdoors and from 47% to 0% for instruction-tuning time backdoors targeting malicious code generation.
Safety backdoor attacks in large language models (LLMs) enable the stealthy triggering of unsafe behaviors while evading detection during normal interactions. The high dimensionality of potential triggers in the token space and the diverse range of malicious behaviors make this a critical challenge. We present BEEAR, a mitigation approach leveraging the insight that backdoor triggers induce relatively uniform drifts in the model's embedding space. Our bi-level optimization method identifies universal embedding perturbations that elicit unwanted behaviors and adjusts the model parameters to reinforce safe behaviors against these perturbations. Experiments show BEEAR reduces the success rate of RLHF time backdoor attacks from >95% to <1% and from 47% to 0% for instruction-tuning time backdoors targeting malicious code generation, without compromising model utility. Requiring only defender-defined safe and unwanted behaviors, BEEAR represents a step towards practical defenses against safety backdoors in LLMs, providing a foundation for further advancements in AI safety and security.