Generating Distributional Adversarial Examples to Evade Statistical Detectors

Yigitcan Kaya, Muhammad Bilal Zafar, Sergül Aydöre, Nathalie Rauschmayr, Krishnaram KenthapadiICML 2022 · OpenReview

Adversarial examples (AEs)—inputs carefully perturbed to fool deep neural networks (DNNs)—have long been the Achilles' heel of modern AI. To counter them, researchers have developed statistical detectors that look for "unnatural" patterns in how these inputs are processed by a model's hidden layers. But are these detectors truly robust, or are they just looking for the wrong things?

A new paper, "Generating Distributional Adversarial Examples to Evade Statistical Detectors," introduces the Statistical Indistinguishability Attack (SIA). This generic adaptive attack proves that if you optimize for it, adversarial examples can be made to look statistically identical to natural data, rendering many current defense strategies ineffective.

The Flaw in Current Defenses

Most adversarial defenses fall into two camps: making the model inherently robust, or detecting AEs before they cause harm. Detection often relies on the assumption that AEs are statistical outliers. For instance, while a natural image of a "cat" might activate certain neurons in a specific way, an adversarial "cat" (which the model misclassifies as a "dog") might produce an anomalous activation pattern that a detector can spot.

However, many of these detectors are evaluated against standard attacks like PGD (Projected Gradient Descent), which only care about changing the model's prediction, not about looking natural. The authors argue this creates a false sense of security. They set out to build a standardized adaptive attack that specifically targets the statistical assumptions these detectors rely on.

Enter SIA: Blending In

The core of SIA is a novel objective function that balances two competing goals:

  1. Attack Success: Using standard cross-entropy loss to ensure the model misclassifies the input.
  2. Statistical Indistinguishability: Using Maximum Mean Discrepancy (MMD) to minimize the distance between the distribution of the adversarial examples and a "guide set" of natural samples.
Standard attacks like PGD and DDN create distinct clusters in the representation space, making them easy to detect compared to evasive methods.

The "secret sauce" of SIA is multi-layer targeting. The authors discovered that an attack optimized to look natural at the penultimate layer might still look like an obvious outlier at earlier layers. SIA solves this by minimizing the MMD loss across all layers of the network simultaneously.

Key Insights and Ideas

  • Single-layer evasion is a leak: If an attacker only targets one layer, a detector operating on a different layer can still catch the anomaly. This highlights the "feature diversity" across DNN layers and proves why multi-layer targeting is essential for a true cloak.
Attacking a single layer often fails to reduce the detection rate at other layers due to feature diversity.
  • Distributional vs. Individual: While SIA is formulated to defeat "distributional detectors" (which analyze a batch of samples), it is surprisingly effective against "individual detectors" (which look at one sample at a time). By following the natural distribution, the AEs naturally fall into high-density regions of the feature space where individual detectors expect natural data to be.
  • Beyond Adversarial Attacks: SIA isn't just for AEs. The authors demonstrated it could also "cloak" corrupted images (like those with snow) or Out-of-Distribution (OOD) samples, making them look like clean, in-distribution data to a detector.

Wake-Up Call Results

The researchers evaluated SIA across several standard benchmarks (CIFAR-10, Tiny-ImageNet) and architectures (ResNet-50, VGG-16, MobileNetV2). The results are a stark warning for the defense community.

Against the state-of-the-art detector SAMMD, which previously boasted near-perfect detection, SIA reduced the detection rate to near 0% while maintaining a ~97-100% attack success rate. In white-box settings, the "Max Layer" detection rate—the absolute best a defender could do across any layer—dropped from 100% for standard attacks to between 3% and 9% for SIA. Even against robust models trained with adversarial training, SIA proved significantly harder to detect than standard PGD.

Limitations and Future Work

While highly powerful, SIA does have limitations:

  • Computational Cost: Computing MMD across all layers makes SIA about 30-40% slower per iteration than standard PGD.
  • Class Distance: The attack is most successful when the starting image and the target class are "close" in representation space (e.g., dog to cat). If they are very far apart (e.g., airplane to frog), the attack has to move the sample further, increasing the risk of detection.
  • Transferability: While SIA evades detectors on the specific model it was crafted for, its "statistical cloak" is less effective when transferred to a completely different, unknown model architecture.

The authors suggest that future work could focus on combining SIA with methods that improve adversarial transferability, potentially creating an even more dangerous "universal" evasive attack. For now, SIA serves as a critical benchmark, proving that statistical detection is far from a solved problem.

Summary updated

Abstract

Deep neural networks (DNNs) are known to be highly vulnerable to adversarial examples (AEs) that include malicious perturbations. Assumptions about the statistical differences between natural and a...
BibTeX
@inproceedings{kaya2022generating,
  author = {Yigitcan Kaya and Muhammad Bilal Zafar and Sergül Aydöre and Nathalie Rauschmayr and Krishnaram Kenthapadi},
  title = {{Generating Distributional Adversarial Examples to Evade Statistical Detectors}},
  booktitle = {International Conference on Machine Learning},
  year = {2022},
  url = {https://openreview.net/forum?id=WLfDtsBSHH7},
}
Figure