Summary of Object-oriented Backdoor Attack Against Image Captioning, by Meiling Li et al.
Object-oriented backdoor attack against image captioning
by Meiling Li, Nan Zhong, Xinpeng Zhang, Zhenxing Qian, Sheng Li
First submitted to arxiv on: 5 Jan 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Artificial Intelligence (cs.AI)
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 This paper explores a new type of attack called a “backdoor” that compromises the performance of vision-language models, specifically image captioning models. The attack involves poisoning training data by modifying pixel values to manipulate the model’s behavior on specific test images. The method designed in this study crafts poisons by considering object regions in an image and adjusts the modification number based on the scale of these regions. After training with poisoned data, the attacked model generates captions that are irrelevant to the given image for specific test images while maintaining performance on benign test images. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Imagine you’re trying to teach a computer to describe pictures by writing sentences about them. Right now, computers can do this pretty well, but someone could secretly change some of the pictures and words to make the computer behave strangely when it sees certain things. This is called a “backdoor” attack. In this study, researchers looked at how to create these backdoors for computers that describe pictures with sentences. They found a way to modify just a few pixels in each picture to make the computer write silly sentences about specific images. But here’s the thing: the computer still does a great job describing normal pictures! This means we need to be careful and find ways to protect computers from these sneaky attacks. |
Keywords
» Artificial intelligence » Image captioning
