Summary of A High Dimensional Statistical Model For Adversarial Training: Geometry and Trade-offs, by Kasimir Tanner et al.
A High Dimensional Statistical Model for Adversarial Training: Geometry and Trade-Offs
by Kasimir Tanner, Matteo Vilucchio, Bruno Loureiro, Florent Krzakala
First submitted to arxiv on: 8 Feb 2024
Categories
- Main: Machine Learning (stat.ML)
- Secondary: Disordered Systems and Neural Networks (cond-mat.dis-nn); Machine Learning (cs.LG)
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 research paper investigates the efficacy of adversarial training in margin-based linear classifiers in high-dimensional spaces where the ratio between data points and dimensionality remains constant. The authors introduce a mathematical model that captures the interplay between data and attacker geometries, allowing for the study of adversarial robustness phenomena. The main theoretical contribution is an exact asymptotic description of sufficient statistics for the adversarial empirical risk minimizer under generic convex and non-increasing losses for Block Feature Models. This allows for the precise characterization of directions in the data associated with a higher generalization/robustness trade-off, as defined by robustness and usefulness metrics. The presence of multiple feature types is crucial to high sample complexity performances of adversarial training, revealing the existence of defendable directions without penalizing accuracy. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This research looks at how well computer programs can work even when they’re given fake or wrong information. They create a special math problem that shows what happens when an attacker tries to trick the program. The results show that some features in the data are more important than others, and that defending those features during training makes the program better at recognizing attacks. |
Keywords
* Artificial intelligence * Generalization
