Poisoned training data for generative LLMs can be filtered by clustering DCT-transformed per-sample gradients, eliminating backdoor attacks without extra retraining.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2024 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Gracefully Filtering Backdoor Samples for Generative Large Language Models without Retraining
Poisoned training data for generative LLMs can be filtered by clustering DCT-transformed per-sample gradients, eliminating backdoor attacks without extra retraining.