KBSOO

Research

Papers and slides, newest first. Select a plate to see it larger.

2

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).

1

HS-JEPA: A JEPA-Inspired Predictive Multi-View Framework for Human Life Logs

Byungsoo Kang, Gwang-Min Park, DongKyu Hyeon (equal contribution)

Workshop on ETRI Human Understanding AI Paper Challenge (IWETRIAI), ICTC 2026, Jeju, Korea, October 2026. Poster.

A prototype that organizes each subject-day of smartphone and wearable life logs into nine label-agnostic semantic views, masks one view, and asks how much of it the rest of the day can explain. Each target is decomposed into a prediction, a residual unexplained by the fitted predictor, and its magnitude (surprise energy). The decomposition is presented as an exploratory analytical framework, not a claim of superior label prediction.