Adversarial Example Detection and Classification With Asymmetrical Adversarial Training

Xuwang Yin, Soheil Kolouri, Gustavo K. Rohde

TL;DR

Researchers have introduced Generative Adversarial Training (GAT), a defense mechanism that decomposes a $K$-class classification problem into $K$ robust binary detectors. By interpreting these detectors as energy-based density models, the system identifies adversarial examples with high precision, ensuring that any successful attack must fundamentally alter the semantic features of the image.

The Vulnerability of Deep Neural Networks

Deep Neural Networks (DNNs) are notoriously brittle. Small, human-imperceptible perturbations—known as adversarial examples—can trick a world-class classifier into misidentifying a stop sign as a speed limit sign.

Existing defenses often fall into two traps:

  1. Heuristic Overfitting: They work against specific attacks but crumble when an attacker employs an "adaptive attack" tailored to the defense.
  2. The "Rubbish" Problem: Discriminative robust classifiers often assign high confidence to "rubbish" images—unrecognizable noise that happens to fall on the wrong side of a decision boundary.

The GAT framework aims to create a principled detection and classification system that remains robust even when the attacker is fully aware of the defense mechanism.

The GAT Framework: One-Versus-the-Rest

The core innovation is a shift from a single $K$-way classifier to a One-Versus-the-Rest (OVR) ensemble of $K$ binary classifiers. For a 10-class problem like MNIST, the system trains 10 separate binary detectors ($d_1, d_2, ... d_{10}$). The $k$-th detector is trained to distinguish between clean samples of class $k$ (positive) and adversarially perturbed samples of all other classes (negative).

Comparison of decision boundaries for a standard classifier versus the proposed robust detectors.

Generative Interpretation

This approach allows for a generative interpretation: the logit output of each binary classifier acts as an unnormalized density model (an energy-based model) of the class-conditional data. Instead of just drawing a line between classes, the model learns the specific distribution of what a "7" or a "dog" actually looks like. This allows the model to reject "rubbish" noise that a standard softmax classifier would be forced to categorize.

Integrated Detection and Classification

The system can operate in two primary modes:

  1. Integrated Detection: A standard classifier provides a label, and the corresponding binary detector "verifies" if the input is a clean sample of that class. If the detector's output is below a threshold $T$, it is flagged as an attack.
  2. Generative Classification: The binary detectors are used directly. The class with the highest logit wins, provided it exceeds the threshold. If all logits are low, the input is rejected.
Detection logic: Adversarial examples are rejected when they fall outside the learned density of the predicted class.

Key Insights

  • Semantic Robustness: In GAT, because the model learns the underlying distribution, a successful attack must actually make the image look like the target class. To fool the "dog" detector, an attacker must synthesize actual dog-like features (ears, fur patterns) onto the original image.
  • Independence of Errors: Because each binary classifier is trained independently on different subspaces, their errors are less likely to be correlated. This makes it significantly harder for an attacker to find a single perturbation that bypasses the entire ensemble.

Evaluation and Results

The authors evaluated GAT on three standard benchmarks: MNIST, CIFAR10, and Restricted ImageNet. The framework demonstrated significant improvements in robustness, particularly against $L_2$ norm-constrained attacks:

  • MNIST: Improved the mean $L_2$ distortion (the amount of change required to fool the model) from 3.68 to 5.65 compared to previous state-of-the-art methods.
  • CIFAR10: Increased mean $L_2$ distortion from 1.1 to 1.5.
  • Interpretability: Qualitative results showed that "attacking" a GAT model to increase a specific class logit actually synthesizes recognizable features of that class, proving the model has learned meaningful semantic features rather than just statistical noise.

Limitations and Future Work

The primary drawback of GAT is computational cost.

  • Inference Speed: Because the model runs $K$ binary checks, inference is roughly $K$ times more expensive than a standard classifier. On CIFAR10, the generative classifier was approximately 10 times slower.
  • Training Complexity: Training $K$ robust detectors requires more memory and time than training a single robust model.

Future research may focus on optimizing the architecture to share parameters between detectors or distilling this generative knowledge into a more efficient single-pass network.

Summary updated

BibTeX
@inproceedings{yin2020adversarial,
  author = {Xuwang Yin and Soheil Kolouri and Gustavo K. Rohde},
  title = {{Adversarial Example Detection and Classification With Asymmetrical Adversarial Training}},
  booktitle = {International Conference on Learning Representations},
  year = {2020},
  url = {https://openreview.net/forum?id=Meyqb3dcKm},
}
Figure