Summary of On the Effectiveness Of Distillation in Mitigating Backdoors in Pre-trained Encoder, by Tingxu Han et al.
On the Effectiveness of Distillation in Mitigating Backdoors in Pre-trained Encoder
by Tingxu Han, Shenghan Huang, Ziqi Ding, Weisong Sun, Yebo Feng, Chunrong Fang, Jun Li, Hanwei Qian, Cong Wu, Quanjun Zhang, Yang Liu, Zhenyu Chen
First submitted to arxiv on: 6 Mar 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: None
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 investigates a defense mechanism against poisoned encoders in self-supervised learning (SSL) called distillation. Initially used in supervised learning to transfer knowledge from a teacher net to a student net, distillation is repurposed here to purify pre-trained encoders tainted by backdoor attacks. The authors conduct an empirical study on the effectiveness of distillation against these attacks, using two state-of-the-art methods and four image classification datasets. Results show that distillation can significantly reduce attack success rates while incurring only a minor loss in accuracy. The paper also delves into the impact of three core components: teacher net, student net, and distillation loss. By exploring 4 teacher nets, 3 student nets, and 6 distillation losses, the authors identify fine-tuned teacher nets, warm-up-training-based student nets, and attention-based distillation loss as the most effective combinations. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper talks about a way to protect computer models from being tricked by bad data. These models are used for things like recognizing pictures. Some people try to make these models do things they don’t want them to, but this new method can help stop that from happening. The researchers tested their idea using some well-known datasets and found it works pretty well. They also looked at how different parts of the system work together to get the best results. |
Keywords
* Artificial intelligence * Attention * Distillation * Image classification * Self supervised * Supervised
