Post-hoc Ranking of Poisoned Training Samples in LoRA-Fine-Tuned Models
LoRA 기반 파인튜닝 모델의 오염 학습 샘플 사후 순위화 기법
Byungsoo Kang, Minseok Kim, Hyungjoon Koo
Conference on Information Security and Cryptography Winter 2026 (CISC-W'26), Yeosu, Korea, October 2026. Oral presentation.
Ranks the training samples of a LoRA-fine-tuned model by suspicion of poisoning. The principal direction of the adapter's weight update is taken by singular value decomposition, and each sample is scored by projecting the hidden representation of its instruction onto that direction, with one forward pass per sample and no additional training or backpropagation. Across 18 settings built from Qwen3-1.7B and BackdoorLLM, the method reached a mean AUROC of 0.857 and a mean average precision of 0.498, higher on average than the reimplemented baselines (Spectral Signatures, SPECTRE, and LoRAScan).