Summary of Badcm: Invisible Backdoor Attack Against Cross-modal Learning, by Zheng Zhang et al.
BadCM: Invisible Backdoor Attack Against Cross-Modal Learning
by Zheng Zhang, Xu Yuan, Lei Zhu, Jingkuan Song, Liqiang Nie
First submitted to arxiv on: 3 Oct 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Cryptography and Security (cs.CR); Machine Learning (cs.LG); Multimedia (cs.MM)
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 introduces a novel bilateral backdoor to address the lack of stealthiness and generalizability in existing cross-modal backdoors. The proposed framework, called BadCM, leverages a cross-modal mining scheme to identify modality-invariant components that can be targeted for poisoning. Well-designed trigger patterns injected into these regions enable efficient recognition by victim models. The framework is adapted to different image-text cross-modal models, making it applicable to various attack scenarios. To generate high-stealthiness poisoned samples, modality-specific generators are developed for visual and linguistic modalities, allowing hiding of explicit trigger patterns in modality-invariant regions. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary BadCM is a new way to make computers misbehave when they’re learning from pictures and words together. Right now, it’s hard to trick these computer systems into making mistakes because they’re not very good at combining information from different types of data (like images and text). The authors created a special kind of “poison” that can be added to the training data to make the computers learn something that’s not useful or correct. This “poison” is hidden in a way that makes it hard for other systems to detect, so even if someone tries to fix the problem, the bad behavior will still happen. |
