Summary of Shifted Interpolation For Differential Privacy, by Jinho Bok et al.
Shifted Interpolation for Differential Privacy
by Jinho Bok, Weijie Su, Jason M. Altschuler
First submitted to arxiv on: 1 Mar 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Cryptography and Security (cs.CR); Optimization and Control (math.OC); Statistics Theory (math.ST); Machine Learning (stat.ML)
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 improves upon previous analyses of noisy gradient descent algorithms by establishing a “privacy amplification by iteration” phenomenon in the framework of f-differential privacy. This unifies all aspects of privacy loss, implying tighter privacy accounting in other notions of differential privacy. The key insight is the construction of shifted interpolated processes that generalize divergence-based relaxations of DP. This leads to exact privacy analysis in strongly convex optimization and extends to various settings, including convex, constrained, full, cyclic, and stochastic batches. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Noisy gradient descent algorithms are used for differentially private machine learning, but it’s hard to quantify how much privacy they leak. This paper helps by showing that as you do more iterations, your data gets more protected from being identified. They use a special framework called f-differential privacy, which is useful because it includes all the ways to measure privacy loss. The main idea is to create new processes that help analyze how much privacy you’re getting with each iteration. This lets them figure out exactly how private your data is in some cases, and they can even apply this to lots of different situations. |
Keywords
* Artificial intelligence * Gradient descent * Machine learning * Optimization
