BEEAR: Embedding-based Adversarial Removal of Safety Backdoors in Instruction-tuned Language Models

arXiv:2406.17092v1 Announce Type: new
Abstract: 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 …

adversarial arxiv attacks backdoor backdoor attacks backdoors behaviors challenge critical cs.ai cs.cr detection enable high language language models large llms malicious normal safety space token