arXiv subjects

Mayuri Sridhar

Publications and source records attributed to Mayuri Sridhar.

5 recordsLinked to original sources

arXiv2026

Sequential stratified inference for the mean

We develop conservative tests for the mean of a bounded population under stratified sampling and apply them to risk-limiting post-election audits. The tests are "anytime valid" under sequential sampling, allowing optional stopping in each stratum. Our core method expresses a global hypothesis about the population mean as a union of intersection hypotheses describing within-stratum means. It tests each intersection hypothesis using a test supermartingale (TSM) combined across strata by multiplication. A $P$-value for each intersection hypothesis is the reciprocal of that TSM, and the largest $P$-value in the union is a $P$-value for the global hypothesis. This approach has two primary moving parts: the rule selecting which stratum to draw from next, and the form of the TSM within each stratum. These rules may vary over intersection hypotheses. We construct intersection tests with the smallest expected sample size and propagate them to approximate an efficient global test. In instances that arise in auditing and other applications, its expected sample size is substantially smaller than that of previous methods.

stat.ME↗

arXiv2026

Private Prediction via PAC Privacy

Machine learning models are increasingly served behind APIs. This renders private prediction, i.e., privatizing a model's outputs rather than its parameters, a natural privacy target: model outputs are lower-dimensional and far more stable to training-data changes than weights. While differential privacy (DP) cannot effectively exploit this as it calibrates noise to worst-case sensitivity that is intractable to bound for non-convex models, we argue that PAC privacy is a natural fit for private prediction. It is instance-based, and calibrates noise to a black-box function's empirical stability to control mutual-information (MI) leakage. The missing ingredient is efficient, adaptive composition. Serving predictions means answering a long stream of adaptively chosen queries from untrusted users; existing composition either fails under adaptivity, grows quadratically, or reverts to input-independent, DP-like noise. We close this gap with a new adversarial composition result via adaptive noise calibration and prove that MI accumulates only linearly under adaptive and adversarial querying. Experiments across modalities show that prediction stability enables high utility even at a tiny per-query budget: on CIFAR-10, we achieve 87.79% accuracy with a per-query MI budget of $2^{-32}$. This enables serving one million queries while provably bounding membership-inference success to 51.08% -- the same guarantee as $(0.04, 10^{-5})$-DP. Further, in the presence of auxiliary public data, the large volume of PAC-private predictions enables us to distill a publishable model that can be queried without limit. Concretely, 210,000 private labels on an ImageNet subset distill into a student reaching 91.86% accuracy on CIFAR-10 with membership inference success bounded by 50.49%, comparable to $(0.02, 10^{-5})$-DP.

cs.LG↗

arXiv2025

Delete and Retain: Efficient Unlearning for Document Classification

Machine unlearning aims to efficiently remove the influence of specific training data from a model without full retraining. While much progress has been made in unlearning for LLMs, document classification models remain relatively understudied. In this paper, we study class-level unlearning for document classifiers and present Hessian Reassignment, a two-step, model-agnostic solution. First, we perform a single influence-style update that subtracts the contribution of all training points from the target class by solving a Hessian-vector system with conjugate gradients, requiring only gradient and Hessian-vector products. Second, in contrast to common unlearning baselines that randomly reclassify deleted-class samples, we enforce a decision-space guarantee via Top-1 classification. On standard text benchmarks, Hessian Reassignment achieves retained-class accuracy close to full retrain-without-class while running orders of magnitude faster. Additionally, it consistently lowers membership-inference advantage on the removed class, measured with pooled multi-shadow attacks. These results demonstrate a practical, principled path to efficient class unlearning in document classification.

cs.LG↗

arXiv2019

$k$-Cut: A Simple Approximately-Uniform Method for Sampling Ballots in Post-Election Audits

We present an approximate sampling framework and discuss how risk-limiting audits can compensate for these approximations, while maintaining their "risk-limiting" properties. Our framework is general and can compensate for counting mistakes made during audits. Moreover, we present and analyze a simple approximate sampling method,"$k$-cut", for picking a ballot randomly from a stack, without counting. Our method involves doing $k$ "cuts", each involving moving a random portion of ballots from the top to the bottom of the stack, and then picking the ballot on top. Unlike conventional methods of picking a ballot at random, $k$-cut does not require identification numbers on the ballots or counting many ballots per draw. We analyze how close the distribution of chosen ballots is to the uniform distribution, and design different mitigation procedures. We show that $k=6$ cuts is enough for an risk-limiting election audit, based on empirical data, which would provide a significant increase in efficiency.

cs.DS↗

arXiv2018

A Peculiarity in the Parity of Primes

We create a simple test for distinguishing between sets of primes and random numbers using just the sum-of-digits function. We find that the sum-of-the-digits of prime numbers does not have an equal probability of being odd or even. The authors know of no reason why prime numbers should bias themselves towards a particular parity in their sums of digits, but our empirical tests show a very strong bias; strong enough that we are able to devise a test to reliably differentiate between collections of prime numbers versus random numbers by looking only at their sums of digits. We are also able to create similar tests for products of primes. We are even able to test "tainted" sets with mixtures of primes and random numbers: as the percentage of (randomly chosen) prime numbers in a set of random numbers is varied, we get a reliable, linear change in our parity measure. For example, when we add up the digits of prime numbers in base 10, their sum is significantly more likely to be odd than even. This effect persists across base changes, although which parity is more common might change. Note that the last digit being odd in base 10 simply reverses the parity. We have tested this for the first fifty million primes -- not primes up to 50,000,000, but the first 50,000,000 prime numbers -- and have found that this effect persists, and does so in a predictable manner. The effect is quite significant; for 50,000,000 primes in base 10, the number of primes which have an odd sum-of-digits is about an order of magnitude farther away from the mean than expected. We have run multiple tests to try and understand the source of this bias, including investigating primes modulo random numbers and adjusting for Chebyshev's bias. None of these tests yielded any satisfactory explanation for this phenomenon.

math.GM↗