Summary of Trasce: Trajectory Steering For Concept Erasure, by Anubhav Jain et al.
TraSCE: Trajectory Steering for Concept Erasure
by Anubhav Jain, Yuya Kobayashi, Takashi Shibuya, Yuhta Takida, Nasir Memon, Julian Togelius, Yuki Mitsufuji
First submitted to arxiv on: 10 Dec 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary A recent paper proposes TraSCE, an approach to guide text-to-image diffusion models away from generating harmful content, such as not-safe-for-work (NSFW) images. Current methods rely on negative prompting, but a widely used strategy can be bypassed in certain cases. The proposed method utilizes a specific formulation of negative prompting and introduces localized loss-based guidance to enhance the technique. This approach achieves state-of-the-art results on various benchmarks, including red team-proposed ones, without requiring training or data modifications. The proposed method is easier for model owners to implement, allowing them to erase new concepts. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Imagine having a special filter that can remove unwanted things from images created by AI models. That’s what the researchers in this paper are working on. They want to help keep these AI models from creating harmful or offensive content. The problem is that some people have found ways to “trick” the models into showing them things they shouldn’t be seeing. To fix this, the researchers came up with a new way to guide the AI models away from creating bad content. This method doesn’t require any special training or data, making it easy for model owners to use. |
Keywords
» Artificial intelligence » Diffusion » Prompting
