Timezone: »
Poster
The advantages of multiple classes for reducing overfitting from test set reuse
Vitaly Feldman · Roy Frostig · Moritz Hardt
Excessive reuse of holdout data can lead to overfitting. However, there is little concrete evidence of significant overfitting due to holdout reuse in popular multiclass benchmarks today.
Known results show that, in the worst-case, revealing the accuracy of $k$ adaptively chosen classifiers on a data set of size $n$ allows to create a classifier with bias of $\Theta(\sqrt{k/n})$ for any binary prediction problem. We show a new upper bound of $\tilde O(\max\{\sqrt{k\log(n)/(mn)}, k/n\})$ on the worst-case bias that any attack can achieve in a prediction problem with $m$ classes. Moreover, we present an efficient attack that achieve a bias of $\Omega(\sqrt{k/(m^2 n)})$ and improves on previous work for the binary setting ($m=2$). We also present an inefficient attack that achieves a bias of $\tilde\Omega(k/n)$. Complementing our theoretical work, we give new practical attacks to stress-test multiclass benchmarks by aiming to create as large a bias as possible with a given number of queries. Our experiments show that the additional uncertainty of prediction with a large number of classes indeed mitigates the effect of our best attacks.
Author Information
Vitaly Feldman (Google Brain)
Roy Frostig (Google Brain)
Moritz Hardt (Google Brain)
Related Events (a corresponding poster, oral, or spotlight)
-
2019 Oral: The advantages of multiple classes for reducing overfitting from test set reuse »
Wed. Jun 12th 10:00 -- 10:05 PM Room Room 103
More from the Same Authors
-
2021 : Lossless Compression of Efficient Private Local Randomizers »
Vitaly Feldman · Kunal Talwar -
2021 : Hiding Among the Clones: A Simple and Nearly Optimal Analysis of Privacy Amplification by Shuffling »
Vitaly Feldman · Audra McMillan · Kunal Talwar -
2021 : Mean Estimation with User-level Privacy under Data Heterogeneity »
Rachel Cummings · Vitaly Feldman · Audra McMillan · Kunal Talwar -
2021 : When Is Memorization of Irrelevant Training Data Necessary for High-Accuracy Learning? »
Gavin Brown · Mark Bun · Vitaly Feldman · Adam Smith · Kunal Talwar -
2023 : Differentially Private Heavy Hitters using Federated Analytics »
Karan Chadha · Junye Chen · John Duchi · Vitaly Feldman · Hanieh Hashemi · Omid Javidbakht · Audra McMillan · Kunal Talwar -
2023 Poster: Near-Optimal Algorithms for Private Online Optimization in the Realizable Regime »
Hilal Asi · Vitaly Feldman · Tomer Koren · Kunal Talwar -
2022 Poster: Optimal Algorithms for Mean Estimation under Local Differential Privacy »
Hilal Asi · Vitaly Feldman · Kunal Talwar -
2022 Oral: Optimal Algorithms for Mean Estimation under Local Differential Privacy »
Hilal Asi · Vitaly Feldman · Kunal Talwar -
2022 Poster: Private frequency estimation via projective geometry »
Vitaly Feldman · Jelani Nelson · Huy Nguyen · Kunal Talwar -
2022 Spotlight: Private frequency estimation via projective geometry »
Vitaly Feldman · Jelani Nelson · Huy Nguyen · Kunal Talwar -
2021 Poster: Lossless Compression of Efficient Private Local Randomizers »
Vitaly Feldman · Kunal Talwar -
2021 Poster: Private Stochastic Convex Optimization: Optimal Rates in L1 Geometry »
Hilal Asi · Vitaly Feldman · Tomer Koren · Kunal Talwar -
2021 Oral: Private Stochastic Convex Optimization: Optimal Rates in L1 Geometry »
Hilal Asi · Vitaly Feldman · Tomer Koren · Kunal Talwar -
2021 Spotlight: Lossless Compression of Efficient Private Local Randomizers »
Vitaly Feldman · Kunal Talwar