Adversarial Risk via Optimal Transport and Optimal Couplings

被引:14
|
作者
Pydi, Muni Sreenivas [1 ]
Jog, Varun [2 ]
机构
[1] Univ Wisconsin, Dept Elect & Comp Engn, 1415 Johnson Dr, Madison, WI 53706 USA
[2] Univ Cambridge, Dept Pure Math & Math Stat, Cambridge CB3 0WB, England
关键词
Couplings; Standards; Measurement; Kernel; Perturbation methods; Loss measurement; Q measurement; Machine learning; statistical learning; robustness; couplings; information theory; DEEP NEURAL-NETWORKS; ROBUST; ALGORITHMS; GO;
D O I
10.1109/TIT.2021.3100107
中图分类号
TP [自动化技术、计算机技术];
学科分类号
0812 ;
摘要
Modern machine learning algorithms perform poorly on adversarially manipulated data. Adversarial risk quantifies the error of classifiers in adversarial settings; adversarial classifiers minimize adversarial risk. In this paper, we analyze adversarial risk and adversarial classifiers from an optimal transport perspective. We show that the optimal adversarial risk for binary classification with 0-1 loss is determined by an optimal transport cost between the probability distributions of the two classes. We develop optimal transport plans (probabilistic couplings) for univariate distributions such as the normal, the uniform, and the triangular distribution. We also derive optimal adversarial classifiers in these settings. Our analysis leads to algorithm-independent fundamental limits on adversarial risk, which we calculate for several real-world datasets. We extend our results to general loss functions under convexity and smoothness assumptions.
引用
收藏
页码:6031 / 6052
页数:22
相关论文
共 50 条
  • [1] Statistical Optimal Transport via Factored Couplings
    Forrow, Aden
    Huetter, Jan-Christian
    Nitzan, Mor
    Rigollet, Philippe
    Schiebinger, Geoffrey
    Weed, Jonathan
    22ND INTERNATIONAL CONFERENCE ON ARTIFICIAL INTELLIGENCE AND STATISTICS, VOL 89, 2019, 89
  • [2] Many Processors, Little Time: MCMC for Partitions via Optimal Transport Couplings
    Nguyen, Tin D.
    Trippe, Brian L.
    Broderick, Tamara
    INTERNATIONAL CONFERENCE ON ARTIFICIAL INTELLIGENCE AND STATISTICS, VOL 151, 2022, 151
  • [3] Quantifying Distributional Model Risk via Optimal Transport
    Blanchet, Jose
    Murthy, Karthyek
    MATHEMATICS OF OPERATIONS RESEARCH, 2019, 44 (02) : 565 - 600
  • [4] Optimal Transport as a Defense Against Adversarial Attacks
    Bouniot, Quentin
    Audigier, Romaric
    Loesch, Angelique
    2020 25TH INTERNATIONAL CONFERENCE ON PATTERN RECOGNITION (ICPR), 2021, : 5044 - 5051
  • [5] Bridging Optimal Transport and Jacobian Regularization by Optimal Trajectory for Enhanced Adversarial Defense
    Le, Binh M.
    Tariq, Shahroz
    Woo, Simon S.
    arXiv, 2023,
  • [6] Lower Bounds on Adversarial Robustness from Optimal Transport
    Bhagoji, Arjun Nitin
    Cullina, Daniel
    Mittal, Prateek
    ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS 32 (NIPS 2019), 2019, 32
  • [7] Generative Adversarial Networks based on optimal transport: a survey
    Bernard Kamsu-Foguem
    Shester Landry Msouobu Gueuwou
    Cheick Abdoul Kadir A. Kounta
    Artificial Intelligence Review, 2023, 56 : 6723 - 6773
  • [8] The multimarginal optimal transport formulation of adversarial multiclass classification
    Trillos, Nicolas Garcia
    Kim, Jakwang
    Jacobs, Matt
    JOURNAL OF MACHINE LEARNING RESEARCH, 2023, 24
  • [9] Generative Adversarial Networks based on optimal transport: a survey
    Kamsu-Foguem, Bernard
    Msouobu Gueuwou, Shester Landry
    Kounta, Cheick Abdoul Kadir A.
    ARTIFICIAL INTELLIGENCE REVIEW, 2023, 56 (07) : 6723 - 6773
  • [10] Robust Risk Management via Multi-marginal Optimal Transport
    Ennaji, Hamza
    Merigot, Quentin
    Nenna, Luca
    Pass, Brendan
    JOURNAL OF OPTIMIZATION THEORY AND APPLICATIONS, 2024, 202 (02) : 554 - 581