Summary of Privacy Of the Last Iterate in Cyclically-sampled Dp-sgd on Nonconvex Composite Losses, by Weiwei Kong et al.
Privacy of the last iterate in cyclically-sampled DP-SGD on nonconvex composite losses
by Weiwei Kong, Mónica Ribero
First submitted to arxiv on: 7 Jul 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Cryptography and Security (cs.CR); Data Structures and Algorithms (cs.DS); Optimization and Control (math.OC); 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 This paper focuses on improving differentially-private stochastic gradient descent (DP-SGD), a family of algorithms used for training machine learning models while ensuring privacy. The existing DP-SGD methods privatize gradients to generate differentially-private model parameters, but they require strong assumptions that are not met in many implementations. These assumptions include knowing the global sensitivity constant, having a Lipschitz or convex loss function, and randomly sampling input batches. To address this challenge, the paper aims to develop tight DP accounting for the last iterate of the algorithm, which would minimize the noise required while maintaining the same privacy guarantee and potentially increasing model utility. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper is about making sure that machine learning models are kept private when they’re being trained. Right now, there’s a way to do this called differentially-private stochastic gradient descent (DP-SGD). It works by adding some noise to the training process so that it’s harder for others to figure out what specific data was used to train the model. The problem is that most people don’t need to keep the entire training process private, just the final model. But current methods require strong assumptions that aren’t always met in practice. This paper wants to find a way to make sure the last step of the training process is also private while still keeping the noise level low. |
Keywords
» Artificial intelligence » Loss function » Machine learning » Stochastic gradient descent
