Label Leakage Attacks in Machine Unlearning: A Parameter and Inversion-Based Approach

Collection of recent security- and privacy-focused ML/LLM papers highlighting several high-impact threats and defenses. Key findings include: (1) Label leakage in machine unlearning via parameter-difference analytics and model-inversion (white-box gradient and black-box genetic attacks) enabling recovery of forgotten classes; (2) RefineRAG — a highly effective word-level retrieval poisoning framework achieving strong transferability to black-box RAG systems; (3) tool-poisoning attack surface and TRUSTDESC — an automated pipeline to generate implementation-faithful tool descriptions to mitigate