Papers by Tran Huynh
BEEAR: Embedding-based Adversarial Removal of Safety Backdoors in Instruction-tuned Language Models (2024.emnlp-main)
Copied to clipboard
| Challenge: | Safety backdoors in large language models can be triggered while evading detection during normal interactions. |
| Approach: | They propose a bi-level optimization method that uses a key insight: backdoor triggers induce a uniform drift in the model’s embedding space . inner level identifies universal perturbations to the decoder’s embedded spaces that steer the model towards defender-defined unwanted behaviors; outer level fine-tunes the model to reinforce safe behaviors against these perturbations. |
| Outcome: | The proposed mitigation method reduces the success rate of safety backdoor attacks from over 95% to 1% for general harmful behaviors and from 47% to 0% for Sleeper Agents, without compromising the model’s usefulness. |