Rearchitecting Classification Frameworks For Increased Robustness

Varun Chandrasekaran, Brian Tang, Nicolas Papernot, Kassem Fawaz, Somesh Jha, Xi Wu

Abstract

While generalizing well over natural inputs, neural networks are vulnerable to adversarial inputs. Existing defenses against adversarial inputs have largely been detached from the real world. These defenses also come at a cost to accuracy. Fortunately, there are invariances of an object that are its salient features; when we break them it will necessarily change the perception of the object. We find that applying invariants to the classification task makes robustness and accuracy feasible together. Two questions follow: how to extract and model these invariances? and how to design a classification paradigm that leverages these invariances to improve the robustness accuracy trade-off? The remainder of the paper discusses solutions to the aforementioned questions.

Introduction

Deep Neural Networks (DNNs) have achieved remarkable success in various domains, yet they remain notoriously brittle. Small, human-imperceptible perturbations—known as adversarial attacks—can easily flip a model's prediction, such as misidentifying a "Stop Sign" as a "Speed Limit" sign. Traditional defenses, including adversarial training and randomized smoothing, typically rely on $p$-norm balls. These mathematical constructs assume attackers are constrained by pixel-level changes, a proxy that often fails to capture human perception and real-world constraints.

Furthermore, these defenses frequently impose a "robustness tax," where gains in adversarial stability come at the expense of accuracy on natural, unperturbed data. This paper argues for a shift toward defenses grounded in physical reality. By identifying and enforcing semantic "invariances"—features like the octagonal shape of a stop sign that must remain for the object to retain its identity—we can build architectures that are inherently more robust without sacrificing performance.

Methodology: Hierarchical Classification

The core of our proposal is a Hierarchical Classification framework. Rather than employing a single "flat" classifier to distinguish between all possible labels simultaneously, we partition the label space based on identified invariants.

The proposed hierarchical classification framework utilizes robust features and intermediate classifiers to partition the label space.

The process involves three primary components:

  1. Invariance Identification: We utilize domain knowledge (e.g., geometric shapes of road signs) and data-driven clustering of neural network embeddings to define "equivalence classes."
  2. Intermediate Classifiers: These nodes enforce high-level invariances. For instance, a root classifier might distinguish between circular and octagonal signs using robust features like LiDAR point clouds, which are immune to 2D adversarial patches.
  3. Leaf Classifiers: Once an input is funneled into an equivalence class, specialized leaf classifiers perform fine-grained prediction. We find that "renormalizing" the output probabilities of a standard model is often more efficient than retraining leaf classifiers from scratch.

Key Insights

By forcing an adversary to operate within a specific equivalence class, we effectively reduce the "attack surface." An attacker must now work significantly harder to flip a label, as they are constrained by the high-level features enforced by the intermediate classifiers. This "Work Factor" principle suggests that enforcing invariants increases the physical cost of an attack, moving beyond simple pixel manipulation.

Experimental Results

We evaluated our framework across computer vision (MNIST, CIFAR-10), real-world traffic sign recognition (GTSRB/KITTI), and audio speaker identification (LibriSpeech).

Comparative performance of the baseline model versus the proposed hierarchical framework.
  • CIFAR-10: The hierarchical approach demonstrated a significant increase in adversarial accuracy. In some configurations, we observed a 6.69% improvement over the baseline while maintaining or even improving natural accuracy.
  • Road Signs: Integrating LiDAR as a robust feature for the root classifier rendered the system immune to 2D adversarial stickers. The certified robustness radius for circular signs showed marked improvement.
  • Audio: In speaker identification tasks, classifying gender as an invariant provided a 1.92x to 2.07x increase in the robustness certificate.

Conclusion and Future Work

Rearchitecting classification to leverage semantic invariants provides a promising path toward breaking the robustness-accuracy trade-off. While identifying these invariants currently relies on a mix of domain expertise and clustering, future work will focus on automating this process through Neural Architecture Search (NAS) to discover optimal hierarchical structures for any given dataset.

Summary updated

Abstract

While generalizing well over natural inputs, neural networks are vulnerable to adversarial inputs. Existing defenses against adversarial inputs have largely been detached from the real world. These defenses also come at a cost to accuracy. Fortunately, there are invariances of an object that are its salient features; when we break them it will necessarily change the perception of the object. We find that applying invariants to the classification task makes robustness and accuracy feasible together. Two questions follow: how to extract and model these invariances? and how to design a classification paradigm that leverages these invariances to improve the robustness accuracy trade-off? The remainder of the paper discusses solutions to the aformenetioned questions.
BibTeX
@misc{chandrasekaran2019rearchitecting,
  author = {Varun Chandrasekaran and Brian Tang and Nicolas Papernot and Kassem Fawaz and Somesh Jha and Xi Wu},
  title = {{Rearchitecting Classification Frameworks For Increased Robustness}},
  year = {2019},
  eprint = {1905.10900},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1905.10900},
}
Figure