Summary of Convergent Privacy Loss Of Noisy-sgd Without Convexity and Smoothness, by Eli Chien et al.
Convergent Privacy Loss of Noisy-SGD without Convexity and Smoothness
by Eli Chien, Pan Li
First submitted to arxiv on: 1 Oct 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Cryptography and Security (cs.CR)
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 studies the differential privacy guarantee of Noisy-SGD algorithms with hidden states over a bounded domain. Traditional analyses assume all internal states are revealed, leading to divergent R’enyi DP bounds. The authors build upon previous work by Ye & Shokri (2022) and Altschuler & Talwar (2022), which proved convergent bounds for smooth convex losses. This paper provides a convergent R’enyi DP bound for non-convex non-smooth losses, showing that requiring Hölder continuous gradients is sufficient. The analysis relies on improved shifted divergence techniques, including forward Wasserstein distance tracking and H”older reduction lemmas. The results demonstrate the benefits of hidden-state analysis for differential privacy. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper looks at how to make sure algorithms are private when they’re learning from data without telling anyone what’s going on inside. Right now, people assume that all the steps they take while learning will be shared with everyone. But this can actually make it harder to keep things private. The authors want to know if there’s a way to keep things private even if we don’t share everything. They found that if the data has some special properties, like having a continuous gradient, then we can still keep things private without sharing all the steps. |
Keywords
» Artificial intelligence » Tracking
