Reproduce “Safe-CLIP: Removing NSFW Concepts from Vision-and-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
Samuele Poppi, Tobia Poppi, Federico Cocchi, Marcella Cornia, Lorenzo Baraldi, Rita Cucchiara
Safe-CLIP introduces a method to fine-tune vision-and-language models like CLIP, creating an embedding space that reduces sensitivity to Not Safe For Work (NSFW) content while preserving performance on safe inputs. This approach demonstrably lowers the probability of generating or retrieving inappropriate content in downstream applications by over 20 percentage points in some cases.
Key insights
- Directly sanitizing the core vision-and-language embedding space offers a robust and systemic solution to AI safety, influencing all downstream applications built upon it.
- A carefully designed multi-modal loss can simultaneously unlearn inappropriate content linkages while preventing catastrophic forgetting of safe concepts and preserving the model's overall expressive power.
- A novel data generation strategy, leveraging a DPO-aligned LLM, can create high-quality synthetic "toxic" data for safety research with minimal human annotation, sidestepping inherent safety mechanisms.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.