Summary of Replace-then-perturb: Targeted Adversarial Attacks with Visual Reasoning For Vision-language Models, by Jonggyu Jang et al.
Replace-then-Perturb: Targeted Adversarial Attacks With Visual Reasoning for Vision-Language Models
by Jonggyu Jang, Hyeonsu Lyu, Jungyeon Koh, Hyun Jong Yang
First submitted to arxiv on: 1 Nov 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Artificial Intelligence (cs.AI); Image and Video Processing (eess.IV)
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 The paper proposes new targeted adversarial attacks for visual-language models (VLMs) that not only alter VLMs’ output text but also maintain the overall integrity of the original image. The authors develop a novel procedure, Replace-then-Perturb, which involves segmenting the target object in an image using a text-guided model, removing it, and inpainting the empty space with the desired prompt. This approach generates targeted adversarial examples that are more effective than existing methods. Additionally, the paper introduces a contrastive learning-based adversarial loss function, Contrastive-Adv, which further improves the performance of the attacks. The authors demonstrate the effectiveness of their proposed methods through extensive benchmark results. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper is about creating fake images to trick AI models that understand pictures and words. These AI models can be trained to do things like identify objects in a picture or describe what’s happening in a scene. But these models are not perfect, and sometimes they get fooled by fake images designed to make them think something else is happening. The authors of this paper want to create even better fake images that can trick these AI models into thinking the wrong thing. They propose two new methods for creating these fake images: one involves finding a specific object in an image and then removing it, and another uses a special type of training that helps make the fake images more realistic. The authors tested their methods on several benchmarks and found that they were more effective at tricking the AI models than other methods. |
Keywords
» Artificial intelligence » Loss function » Prompt
