[Paper Review] Adversarial Examples for Unsupervised Machine Learning Models.

Chia-Yi Hsu, Pin‐Yu Chen+3 more|arXiv (Cornell University)|Mar 2, 2021

Adversarial Robustness in Machine Learning39 references4 citations

TL;DR

This paper introduces a novel framework for generating adversarial examples in unsupervised machine learning without relying on labels or pre-trained classifiers. By leveraging a mutual information neural estimator as an unsupervised similarity measure and a provably convergent MinMax algorithm, the method enables effective data augmentation that significantly improves performance across unsupervised tasks like representation learning and contrastive learning.

ABSTRACT

Adversarial examples causing evasive predictions are widely used to evaluate and improve the robustness of machine learning models. However, current studies on adversarial examples focus on supervised learning tasks, relying on the ground-truth data label, a targeted objective, or supervision from a trained classifier. In this paper, we propose a framework of generating adversarial examples for unsupervised models and demonstrate novel applications to data augmentation. Our framework exploits a mutual information neural estimator as an information-theoretic similarity measure to generate adversarial examples without supervision. We propose a new MinMax algorithm with provable convergence guarantees for efficient generation of unsupervised adversarial examples. Our framework can also be extended to supervised adversarial examples. When using unsupervised adversarial examples as a simple plug-in data augmentation tool for model retraining, significant improvements are consistently observed across different unsupervised tasks and datasets, including data reconstruction, representation learning, and contrastive learning. Our results show novel methods and advantages in studying and improving robustness of unsupervised learning problems via adversarial examples. Our codes are available at this https URL.

Motivation & Objective

• To address the lack of adversarial example generation methods for unsupervised machine learning models, which currently rely on supervision or labels.
• To develop a label-free, self-supervised method for generating adversarial examples that can improve model robustness in unsupervised settings.
• To demonstrate the utility of unsupervised adversarial examples as a plug-in data augmentation technique for enhancing model performance.
• To provide theoretical convergence guarantees for the proposed MinMax optimization framework in the absence of supervision.
• To extend the applicability of adversarial examples beyond supervised learning to unsupervised representation learning and contrastive learning.

Proposed method

• The framework uses a mutual information neural estimator (MINE) as an information-theoretic similarity measure to guide adversarial example generation without requiring ground-truth labels.
• A new MinMax optimization algorithm is proposed to generate adversarial examples, with theoretical convergence guarantees under the unsupervised setting.
• The method operates by perturbing input data to maximize the mutual information between original and perturbed inputs, thereby creating adversarial examples that challenge model generalization.
• The framework is designed to be modular and can be extended to supervised learning by incorporating label-based objectives.
• Adversarial examples are used as data augmentation by retraining models on the original and adversarial samples, improving robustness and performance.
• The approach is evaluated across multiple unsupervised tasks, including autoencoders, contrastive learning, and representation learning.

Experimental results

Research questions

• RQ1 Can adversarial examples be effectively generated for unsupervised machine learning models without access to ground-truth labels or pre-trained classifiers?
• RQ2 How can mutual information estimation be leveraged as a proxy for adversarial perturbation in the absence of supervision?
• RQ3 What is the impact of using unsupervised adversarial examples as data augmentation on downstream unsupervised learning performance?
• RQ4 Does the proposed MinMax algorithm with provable convergence improve the efficiency and stability of adversarial example generation in unsupervised settings?
• RQ5 Can the framework be extended to supervised learning while maintaining robustness and performance gains?

Key findings

• The proposed method successfully generates adversarial examples for unsupervised models without relying on labels or classifier outputs, enabling robustness evaluation in previously unaddressed settings.
• The use of mutual information neural estimation as a similarity measure enables effective and label-free adversarial example generation in unsupervised learning.
• The MinMax algorithm with provable convergence guarantees ensures stable and efficient generation of adversarial examples in the absence of supervision.
• When applied as data augmentation, unsupervised adversarial examples lead to consistent and significant performance improvements across diverse unsupervised tasks, including representation learning and contrastive learning.
• The framework demonstrates transferability to supervised learning, showing that unsupervised adversarial examples can also enhance performance in supervised settings.
• The method achieves state-of-the-art improvements in reconstruction accuracy and representation quality on benchmark datasets, validating its effectiveness and generalization capability.