Removing the NEEDLE in the Haystack: Backdoor Removal in LLMs via Weight Orthogonalisation
NEEDLE removes LLM backdoors by weight orthogonalisation with little capability or safety loss.
NEEDLE is a training-free LLM backdoor removal method that estimates a backdoor direction and a refusal subspace from activation vectors, then applies sequential weight orthogonalisation. It does not require a clean reference model or the original poisoned training data. Across model families and attack types it reports the lowest mean attack success rate among evaluated defenses, including 0% on code-injection attacks, with the lowest KL divergence and minimal capability and safety change.
- Training-free removal once a trigger is identified
- Orthogonalizes a backdoor direction while preserving the refusal subspace
- Lowest mean ASR among compared defenses, including 0% on code injection
- Lowest KL divergence with minimal capability and safety change
From the original article
1 min read at huggingface.co
Backdoor attacks can be implanted in Large Language Models (LLMs) during training, causing unwanted behaviour when a trigger appears in the input. Existing backdoor defences for LLMs attempt to remove the backdoor but inadvertently shift the model's output distribution to benign prompts, which can result in degraded model performance and safety. We propose NEEDLE, a training-free method for targeted backdoor removal.
By Minoo Kim, Vasileios Lampos, George Drayson · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.