Learning Adversarially Robust Representations via Worst-Case Mutual Information Maximization

Sicheng Zhu, Xiao Zhang, David Evans

In the quest for adversarial robustness, most research focuses on fortifying the final classifier. This paper, presented at ICML 2020, shifts the focus to the underlying representations the model learns. By defining a new metric called "representation vulnerability" based on mutual information, the authors provide both a theoretical framework for understanding why models fail and an unsupervised method to train more resilient features.

The Problem: Brittle Features Lead to Brittle Models

Despite the success of deep learning, models remain notoriously "brittle"—small, imperceptible perturbations (adversarial attacks) can cause them to fail spectacularly. Most existing defenses, like adversarial training, focus on the end-to-end classification task. However, recent theory suggests that robustness might be intrinsically difficult due to the geometry of high-dimensional data.

The authors argue that we should look at the representations (the features extracted by the middle layers of a network). If the features themselves are highly sensitive to small changes in the input, no amount of clever classification logic can fix the resulting errors. The challenge is defining what a "robust representation" actually is without relying on labels, and then finding a way to train for it.

Methods: Quantifying and Minimizing Vulnerability

The paper introduces two core concepts to address this challenge:

  1. Representation Vulnerability (RV): This metric is defined as the maximum change in Mutual Information (MI) between the input and the representation when the input is perturbed within an $\infty$-Wasserstein ball (representing small, worst-case changes). If a representation is robust, the information it carries about the input shouldn't fluctuate wildly when the input is slightly nudged.
  2. Worst-Case Mutual Information Maximization: Standard representation learning often tries to maximize the MI between the input $X$ and the representation $Z$. The authors propose a "min-max" objective: $$\max_{g} \min_{\mu_{X'}} I(X'; g(X'))$$ In plain English: Train the encoder ($g$) to maximize the information it retains, even when an adversary is trying to perturb the input distribution ($\mu_{X'}$) to minimize that information.

To implement this, they use the Mutual Information Neural Estimator (MINE) to approximate the MI and a projected gradient descent (PGD) approach to find the worst-case perturbations during training. Crucially, this process is unsupervised—it doesn't require labels to learn robust features.

Key Insights and Ideas

  • The Information-Theoretic Bottleneck: The paper establishes a theoretical lower bound on adversarial risk. It proves that if a representation has high vulnerability or low initial mutual information, it is mathematically impossible to build a robust classifier on top of it.
  • Robustness is Unsupervised: One of the most surprising findings is that you can induce adversarial robustness without ever looking at a label. By simply forcing the encoder to maintain stable information flow under perturbation, the resulting features are naturally better suited for robust classification.
  • Alignment with Human Perception: The authors found that robust representations lead to much cleaner saliency maps. While standard models often focus on pixel-level statistical noise, robust representations focus on higher-level visual concepts (shapes and structures) that align better with human perception.

Key Results

The paper empirically confirms that lower Representation Vulnerability directly correlates with a smaller "adversarial gap" (the difference between natural and adversarial accuracy).

Correlation between Representation Vulnerability and the adversarial accuracy gap.

The authors evaluated their approach on standard computer vision benchmarks (MNIST, Fashion-MNIST, SVHN, and CIFAR-10). When using their unsupervised robust representations combined with a robust downstream classifier, they achieved significant improvements over standard unsupervised baselines.

Performance comparison showing the proposed method achieving 31.52% adversarial accuracy.

As shown in the table above, the proposed method achieves 31.52% adversarial accuracy on CIFAR-10. For comparison, a standard unsupervised representation (like Hjelm et al., 2018) achieves nearly 0% under the same attack when paired with a standard classifier. Furthermore, their theoretical lower bound on adversarial risk was shown to be reasonably tight, providing a useful practical estimate of the best possible robustness a specific architecture can achieve.

Limitations and Future Work

While the unsupervised method is a huge leap over standard unsupervised learning, it still trails behind fully supervised adversarial training (which hits ~40.5% on the same setup). Additionally, estimating mutual information using MINE and performing inner-loop optimization for worst-case perturbations is computationally expensive. Finally, scaling this unsupervised principle to larger, more complex datasets like ImageNet remains an open challenge, and the choice of the downstream classifier still matters significantly to fully realize the representation's robustness.

Summary updated

Abstract

Training machine learning models that are robust against adversarial inputs poses seemingly insurmountable challenges. To better understand adversarial robustness, we consider the underlying proble...
BibTeX
@inproceedings{zhu2020learning,
  author = {Sicheng Zhu and Xiao Zhang and David Evans},
  title = {{Learning Adversarially Robust Representations via Worst-Case Mutual Information Maximization}},
  booktitle = {International Conference on Machine Learning},
  year = {2020},
  url = {https://openreview.net/forum?id=5-hHA9yqQmT},
}
Figure