Mitigating Information Leakage in Image Representations: A Maximum\n Entropy Approach

2019/04/10 by Proteek Chandan Roy, Vishnu Naresh Boddeti, Roy, Proteek Chandan +1 · 8 citations
Computer Science · Engineering · #Adversarial Robustness in Machine Learning #Computer Vision and Pattern Recognition (cs.CV) #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning (stat.ML) #Sparse and Compressive Sensing Techniques

paper · pdf · doi:10.48550/arxiv.1904.05514

openalex publication_date 2019/04/10 · openalex created_date 2022/07/29 · openalex updated_date 2026/07/28

Abstract

Image recognition systems have demonstrated tremendous progress over the past\nfew decades thanks, in part, to our ability of learning compact and robust\nrepresentations of images. As we witness the wide spread adoption of these\nsystems, it is imperative to consider the problem of unintended leakage of\ninformation from an image representation, which might compromise the privacy of\nthe data owner. This paper investigates the problem of learning an image\nrepresentation that minimizes such leakage of user information. We formulate\nthe problem as an adversarial non-zero sum game of finding a good embedding\nfunction with two competing goals: to retain as much task dependent\ndiscriminative image information as possible, while simultaneously minimizing\nthe amount of information, as measured by entropy, about other sensitive\nattributes of the user. We analyze the stability and convergence dynamics of\nthe proposed formulation using tools from non-linear systems theory and compare\nto that of the corresponding adversarial zero-sum game formulation that\noptimizes likelihood as a measure of information content. Numerical experiments\non UCI, Extended Yale B, CIFAR-10 and CIFAR-100 datasets indicate that our\nproposed approach is able to learn image representations that exhibit high task\nperformance while mitigating leakage of predefined sensitive information.\n

Cited by

• RMLer: Synthesizing Novel Objects across Diverse Categories via Reinforcement Mixing Learning
• InfoDecom: Decomposing Information for Defending Against Privacy Leakage in Split Inference
• Fair Representation Learning with Controllable High Confidence Guarantees via Adversarial Inference
• A Multi-Objective Evaluation Framework for Analyzing Utility-Fairness Trade-Offs in Machine Learning Systems
• Learning Private Representations through Entropy-based Adversarial Training
• Fair Domain Generalization: An Information-Theoretic View
• PASS: Private Attributes Protection with Stochastic Data Substitution
• Simple and Effective Specialized Representations for Fair Classifiers

Related

• Learning Private Representations through Entropy-based Adversarial Training
• Robust Machine Learning via Privacy/Rate-Distortion Theory
• Learning Privacy Preserving Encodings through Adversarial Training
• InfoScrub: Towards Attribute Privacy by Targeted Obfuscation
• Adversarially Robust AI-Generated Image Detection for Free: An Information Theoretic Perspective