Researchers Propose Expanded Benchmarks for Differentially Private Image Classification

Global: Researchers Propose Expanded Benchmarks for Differentially Private Image Classification

Researchers from multiple institutions announced a revised suite of benchmarks aimed at evaluating differentially private image classification models. The work was first submitted to arXiv on 23 January 2026 and revised on 27 January 2026. By broadening evaluation scenarios, the authors seek to provide a more comprehensive picture of privacy‑preserving learning performance.

Expanded Benchmark Suite

The proposed benchmarks cover a range of conditions, including experiments with and without auxiliary data, assessments in convex optimization settings, and tests across several qualitatively distinct image datasets. This variety is intended to reflect real‑world deployment contexts and to challenge methods beyond narrow, idealized scenarios.

Evaluation of Existing Techniques

Using the new benchmark collection, the authors re‑examined established differentially private training algorithms. Their results indicate that some previously reported advantages persist across multiple settings, while other techniques show diminished effectiveness when evaluated under the broader conditions.

Community Leaderboard

To facilitate ongoing progress, the team created a publicly accessible leaderboard where researchers can submit results and compare performance against the expanded benchmarks. The platform is designed to track advancements in privacy‑preserving image classification over time.

Publication and Citation Details

The paper, titled “Rethinking Benchmarks for Differentially Private Image Classification,” is listed under the arXiv identifier arXiv:2601.17189 [cs.LG] and carries the DOI https://doi.org/10.48550/arXiv.2601.17189. It has also been referenced in the IEEE Technical Committee on Data Engineering 2025 proceedings.

Authors and Availability

Authored by Sabrina Mokhtari, Sara Kodeiri, Shubhankar Mohapatra, Florian Tramèr, and Gautam Kamath, the full manuscript (2,094 KB) is available as a PDF on arXiv. The submission history records two versions, with the latest uploaded on 27 January 2026.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.