Training-time defense that redistributes LLM refusal via mean/covariance matching + KD, raising linear-ablation attack rank from K=1 to K≥16 (Llama-3.2-1B-Instruct)
-
Updated
Jun 12, 2026 - Python
Training-time defense that redistributes LLM refusal via mean/covariance matching + KD, raising linear-ablation attack rank from K=1 to K≥16 (Llama-3.2-1B-Instruct)
Domain-adapted OCR for multilingual medical documents: LoRA/PEFT adaptation of compact VLMs, seed-level reproducibility and an evaluation gate, the OCR error cascade under shortcut control, and redaction mechanisms (masking, adversarial gate, LEACE) under probe and inversion attacks. Paper code + Zenodo data (10.5281/zenodo.22078532).
To associate your repository with the leace topic, visit your repo's landing page and select "manage topics."