Summary of Average Certified Radius Is a Poor Metric For Randomized Smoothing, by Chenhao Sun et al.
Average Certified Radius is a Poor Metric for Randomized Smoothing
by Chenhao Sun, Yuhao Mao, Mark Niklas Müller, Martin Vechev
First submitted to arxiv on: 9 Oct 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 The paper investigates the effectiveness of randomized smoothing in providing certified robustness guarantees against adversarial attacks. It highlights that the average certified radius (ACR) is a popular metric for comparing methods and tracking progress in the field, but shows that ACR is a poor metric for evaluating robustness guarantees provided by randomized smoothing. The authors theoretically prove that a trivial classifier can have arbitrarily large ACR and empirically confirm that existing training strategies improve ACR on easy samples but reduce robustness on hard ones. To strengthen their conclusion, they propose strategies to achieve state-of-the-art ACR without training for robustness on the full data distribution. Overall, the paper suggests that ACR has introduced an undesired bias to the field and its application should be discontinued when evaluating randomized smoothing. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper looks at a way to make computers more secure from attacks. Right now, people are using something called “randomized smoothing” to try to keep their models safe. They’re looking at how well this works by measuring something called the average certified radius (ACR). But the authors of the paper think that ACR isn’t a good way to measure this because it can be tricked into thinking a bad model is actually good if you make it really good on easy problems, even if it’s terrible on hard ones. They show that existing ways of training these models are making them worse at handling hard problems. The authors suggest some new ways to train the models that might help. |
Keywords
» Artificial intelligence » Tracking
