GradSentry: Gradient Spectral Entropy for Backdoor Sample Filtering in Large Language Model Fine-Tuning

Researchers introduced GradSentry, a backdoor defense for LLM fine-tuning that uses spectral entropy of per-sample gradients to spot poisoned samples.

It separates poisoned data from clean data without clustering and works for both LoRA and full-parameter tuning.

GradSentry stays effective across poison ratios from 1% to 90% with low overhead.

The method shows strong results on multiple QA datasets and attack types.

It provides a fast, training-agnostic way to filter poisoned fine-tuning samples, even at extreme poison rates.