Reproduce “Single Image Unlearning: Efficient Machine Unlearning in Multimodal Large Language Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jiaqi Li, Qianshan Wei, Chuanyi Zhang, Guilin Qi, Miaozeng Du, Yongrui Chen, Sheng Bi, Fan Liu

Researchers at Southeast University and Hohai University introduced Single Image Unlearning (SIU), a method enabling efficient visual concept unlearning in Multimodal Large Language Models (MLLMs) using just one image per concept. SIU achieved a 99.0% Exact Match on unseen images for LLAVA-7B, preserving overall model utility, fluency, and diversity, while effectively countering membership inference and jailbreak attacks.

Key insights

  • Effective visual unlearning in MLLMs is achievable with minimal data (a single image) by strategically structuring fine-tuning data to guide both targeted forgetting and general knowledge preservation.
  • The Dual Masked KL-divergence (DMK) Loss successfully resolves the conflict between specific token suppression and general utility preservation, preventing common model degradation issues.
  • Unlearning is not merely an erasure but a complex restructuring of knowledge, allowing for the removal of visual recognition while retaining related factual information, as evidenced by 'positive butterfly effects'.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.