Bar Mahpud

Bar Mahpud

PhD candidate in Computer Engineering at Bar-Ilan University specializing in Differential Privacy and Machine Learning, ex Head of Department at the IDF's J6 & Cyber Defence Directorate.

Publications

You can also find my articles on my Google Scholar profile.

Conference Papers

A Private Approximation of the 2nd-Moment Matrix of Any Subsamplable Input

Published in NeurIPS 2025

We present a zCDP-compliant recursive algorithm for differentially private second-moment estimation that exploits subsamplability assumptions to achieve practical spectral accuracy on worst-case data and remain robust to outliers.

Recommended citation: Mahpud, Bar; Sheffet, Or. (2025). "A Private Approximation of the 2nd-Moment Matrix of Any Subsamplable Input." NeurIPS 2025.
Download Paper | Download Slides | Download Bibtex

Privacy-Preserving Conformal Prediction Under Local Differential Privacy

Published in COPA 2025

We propose two LDP conformal prediction frameworks that compute thresholds from privatized data without true labels—(i) users send features and a k-ary randomized-response label (no model access) and (ii) users privatize conformity via binary-search responses (requires model access, stricter privacy)—and prove finite-sample coverage with robust performance under strong randomization for sensitive applications.

Recommended citation: Penso, Coby; Mahpud, Bar; Goldberger, Jacob; Sheffet, Or. (2025). "Privacy-Preserving Conformal Prediction Under Local Differential Privacy." COPA 2025.
Download Paper | Download Slides | Download Bibtex

Differentially Private Approximation of the John Ellipsoid

Published in (under review)

We develop the first differentially private algorithm for approximating the John ellipsoid under the standard privacy model, using a projected multiplicative-weights framework with rigorous privacy–utility guarantees.

Recommended citation: Mahpud, Bar; Sheffet, Or; Omer, Daniel. (2026). "Differentially Private Approximation of the John Ellipsoid." (under review).
Download Paper | Download Bibtex

A Differentially Private Linear-Time fPTAS for the Minimum Enclosing Ball Problem

Published in NeurIPS 2022

We present the first differentially private FPTAS for the Minimum Enclosing Ball with near-linear runtime, near-optimal radius, and few uncovered points; we also introduce a local-privacy version that improves prior results and show strong empirical performance.

Recommended citation: Mahpud, Bar; Sheffet, Or. (2022). "A Differentially Private Linear-Time fPTAS for the Minimum Enclosing Ball Problem." NeurIPS 2022.
Download Paper | Download Slides | Download Bibtex

Differentially Private Learning of Exponential Distributions: Adaptive Algorithms and Tight Bounds

Published in (under review)

We develop two complementary pure-DP methods for learning Exponential Distributions—a clipped-noisy MLE and a quantile estimator—combine them adaptively to attain near-optimal sample complexity; extend via a logarithmic reduction to Pareto; prove nearly matching lower bounds (packing & group privacy); and show that approximate-DP removes the need for external bounds.

Recommended citation: Mahpud, Bar; Sheffet, Or. (2025). "Differentially Private Learning of Exponential Distributions: Adaptive Algorithms and Tight Bounds." (under review).
Download Paper | Download Bibtex