Toward Understanding Adversarial Distillation: Why Robust Teachers Fail

Hongsin Lee, Hye Won Chung

Introduction

Deep neural networks are notoriously fragile. Small, intentional perturbations—known as adversarial attacks—can easily trick them into making incorrect predictions. Adversarial Training (AT) is the standard defense against these vulnerabilities, but it suffers from a phenomenon known as "robust overfitting," where the model stops generalizing and instead begins memorizing spurious noise.

To mitigate this, researchers frequently employ Adversarial Distillation (AD). In this setup, a larger, robust "teacher" model guides a smaller "student" model using soft labels. Intuitively, a more robust teacher should produce a more robust student. However, empirical evidence reveals a frustrating paradox: sometimes, a state-of-the-art, highly robust teacher actually harms the student's performance. This paper investigates the hidden mechanisms causing high-quality teachers to fail their students and offers a principled solution for teacher selection.

The "Bad" Teacher Paradox

In the context of adversarial distillation, a "bad" teacher is rarely a weak model. Paradoxically, it is often a stronger model that is overly confident. Because the teacher "knows" too much and relies on complex robust features, it provides a supervisory target that the capacity-constrained student physically cannot represent. To minimize the training loss and match the teacher's confident predictions, the student is forced to "cheat" by memorizing spurious noise patterns, ultimately leading to robust overfitting.

The Robustly Unlearnable Set

To understand this dynamic, the authors identify a consistent subset of training data that models simply cannot learn robustly, termed the Robustly Unlearnable Set ($S_U$).

Crucially, unlearnability is relative. A sample is not inherently "unlearnable" because it is corrupted; rather, it is unlearnable because the student lacks the architectural capacity to extract the specific features the teacher is utilizing. The authors found that this set exists across all tested architectures (from MobileNet-V2 to WideResNet-34-10). As model capacity increases, the size of this unlearnable set naturally shrinks (e.g., from approximately 9,000 down to 1,500 samples in CIFAR-10).

Theoretical Framework: The Distillation Dichotomy

The researchers developed a rigorous theoretical framework analyzing the feature learning dynamics of a two-layer neural network. They modeled a scenario where a teacher has access to complex robust features that the student cannot represent, mathematically proving a clear dichotomy in distillation outcomes:

  • The Confident Teacher (Overfitting): If a teacher provides high-confidence labels on the unlearnable samples ($S_U$), the student is compelled to memorize spurious noise patterns. This noise eventually overpowers the learned robust signal, driving robust overfitting.
  • The Uncertain Teacher (Generalization): Conversely, if a teacher exhibits high predictive entropy (uncertainty) on $S_U$, it acts as an implicit regularizer. By signaling uncertainty, the teacher effectively tells the student not to over-fit these difficult samples. This suppresses noise memorization and allows the student to rely solely on learnable signals for robust generalization.

Empirical Validation and Key Results

The theoretical findings were validated across synthetic simulations and real-image classification benchmarks, including CIFAR-10, CIFAR-100, and Tiny-ImageNet, utilizing a wide array of pre-trained robust models from RobustBench.

The results were striking:

  • Overfitting Mitigation: Utilizing a "High Entropy" teacher drastically reduced robust performance degradation (the drop from peak to final accuracy) from an average of 7.23% down to just 0.53%.
  • Predictive Power: A teacher's overall robust accuracy proved to be a poor predictor of distillation success. However, the teacher's entropy on the unlearnable set correlated almost perfectly (Pearson $r > 0.9$) with the student's final robustness.

A Practical Solution: The Unlearnable-Entropy Criterion

Based on these insights, the authors propose the Unlearnable-Entropy Criterion (UEC) as a practical selection rule. Instead of blindly choosing the teacher with the highest robust accuracy, practitioners should select teachers that exhibit high predictive entropy on a proxy unlearnable set (which can be identified using a standard baseline model).

The UEC successfully identified high-quality teachers across five different distillation algorithms (including ARD, IAD, and RSLAD), demonstrating that a teacher's behavior on $S_U$ is far more critical than the specific distillation loss function employed.

Limitations and Future Directions

While this research provides a significant breakthrough in understanding adversarial distillation, several avenues remain open for future exploration:

  • Architectural Complexity: The theoretical proofs focus on a two-layer network. While empirical results on deeper architectures like ResNets are strong, the mathematical dynamics of deeper networks are inherently more complex.
  • Proxy Dependency: The UEC relies on a "proxy" unlearnable set generated by a standard PGD-AT model. While efficient, the "gold standard" for identifying $S_U$ involves the intersection of multiple models, which is computationally expensive.
  • Beyond Image Classification: The current study is limited to image classification. How these "unlearnable sets" and capacity mismatches manifest in large language models (LLMs) or generative tasks remains an exciting open question.

Summary updated

Abstract

Adversarial Distillation aims to enhance student robustness by guiding the student with a robust teacher's soft labels within the min-max adversarial training framework, yet its success is notoriously inconsistent: a more robust teacher often fails to improve, or even harms, the student's robust generalization. In this paper, we identify a key mechanism of this teacher dependency: the misalignment between the teacher's supervisory confidence and the student's representational limitations on a consistent subset of training data—the Robustly Unlearnable Set. We present a theoretical framework analyzing the feature learning dynamics of a two-layer neural network, demonstrating that this mismatch creates a dichotomy in distillation outcomes. We prove that when a teacher provides confident supervision on unlearnable samples, it compels the student to memorize spurious noise patterns that eventually overpower the learned robust signal, thereby driving robust overfitting. Conversely, a teacher that exhibits high uncertainty on these samples effectively suppresses noise memorization, allowing the student to rely solely on the learnable signal for robust generalization. We empirically validate our theory across both synthetic simulations and real-image classification datasets, confirming that robust overfitting is driven by the teacher's interaction with unlearnable samples. Finally, we demonstrate that a teacher's predictive entropy on unlearnable samples serves as a strong indicator of student robustness, validating our theoretical framework and offering a principled guideline for robust teacher selection.
BibTeX
@inproceedings{lee2026toward,
  author = {Hongsin Lee and Hye Won Chung},
  title = {{Toward Understanding Adversarial Distillation: Why Robust Teachers Fail}},
  booktitle = {International Conference on Machine Learning},
  year = {2026},
  url = {https://openreview.net/forum?id=USejwpvj0a},
}
Figure