Abstract
The problem of adversarial examples has shown that modern Neural Network (NN) models could be rather fragile. Among the more established techniques to solve the problem, one is to require the model to be $\epsilon$-adversarially robust (AR); that is, to require the model not to change predicted labels when any given input examples are perturbed within a certain range. However, it is observed that such methods lead to standard performance degradation, i.e., the degradation on natural examples. In this work, we study the degradation through the regularization perspective. We identify quantities from generalization analysis of NNs; with the identified quantities we empirically find that AR is achieved by regularizing/biasing NNs towards less confident solutions by making the changes in the feature space (induced by changes in the instance space) of most layers smoother uniformly in all directions; so to a certain extent, it prevents sudden change in prediction w.r.t. perturbations. However, the end result of such smoothing concentrates samples around decision boundaries, resulting in less confident solutions, and leads to worse standard performance. Our studies suggest that one might consider ways that build AR into NNs in a gentler way to avoid the problematic regularization.
Introduction: The Robustness-Accuracy Trade-off
Deep Neural Networks (NNs) are notoriously fragile; tiny, imperceptible perturbations, known as adversarial examples, can easily fool them. The standard defense is Adversarial Training (AT), which forces the model to be $\epsilon$-adversarially robust. However, researchers have long observed a frustrating trade-off: as robustness increases, standard accuracy on "natural" images decreases.
This paper investigates this phenomenon through the lens of regularization. The authors ask: How does the requirement of robustness change the internal behavior of the network? Does it act like a standard regularizer (like weight decay), or is it fundamentally different?
The Geometry of Robustness
To understand the trade-off, we must consider the relationship between decision boundaries and the "margin"—the distance from data points to these boundaries. Adversarial robustness requires that no point within an $\epsilon$-radius of an input changes its predicted label.
As illustrated in the figure above, enforcing a robustness radius $\epsilon$ effectively "weeds out" certain hyperplanes from the hypothesis space. This forces the decision boundary to maintain a minimum distance from the training samples, which provides stability but limits the model's flexibility in fitting the data perfectly.
Methods and Theoretical Framework
The authors combine theoretical derivation with extensive empirical analysis on benchmarks like CIFAR-10, CIFAR-100, and Tiny-ImageNet. They establish a new upper bound for the Generalization Error (GE) of adversarially robust NNs. This bound explicitly links the generalization gap to two key quantities:
- Instance-space margin: The distance to decision boundaries.
- Singular values: The values in weight matrices that determine how much a layer amplifies or shrinks an input.
Empirical Evidence of Regularization
By tracking these quantities across different levels of robustness strength (measured by the $\epsilon$ perturbation allowed), the authors reveal how AR acts as a "clumsy" regularizer.
1. The Singular Value "Squeeze"
Robustness is achieved by making the mapping from input to feature space "smoother" in all directions. Empirically, the authors found that AR reduces the standard deviation of singular values in weight matrices (Figure 4a). By making these values more uniform, the network avoids amplifying noise in specific directions, which prevents small input changes from causing large output swings.
2. The Confidence Crisis
This smoothing has a significant side effect: it concentrates samples around the decision boundaries. In the plots of margin distributions (Figure 4b), as robustness strength increases from 4 to 16, the peak of the distribution moves toward zero.
Because the model is forced to be smooth, it becomes "less confident." Interestingly, on datasets like CIFAR-100, while the error rate goes up, the test loss sometimes goes down. This is because the model's mistakes are "gentler"—it assigns lower probability to its wrong guesses than a non-robust model would. However, this indecision ultimately hurts standard test accuracy (Figure 4c).
Conclusion
The study concludes that current AR methods are problematic because they smooth out feature changes uniformly. This makes the network unable to distinguish between adversarial noise and actual differences between classes. The findings suggest a need for "gentler" ways to build robustness into NNs—methods that provide stability without forcing the model into a state of low-confidence indecision.
Summary updated