Enhancing ML Robustness Using Physical-World Constraints

Varun Chandrasekaran, Brian Tang, Varsha Pendyala, Kassem Fawaz, Somesh Jha, Xi Wu

Introduction

Deep Neural Networks (DNNs) have revolutionized computer vision and speech recognition, yet they remain notoriously fragile. Small, often imperceptible perturbations—known as adversarial attacks—can cause these models to fail catastrophically. Traditional defenses typically rely on mathematical constraints, such as ensuring output stability within a small "p-norm ball." However, these methods often suffer from a significant "robustness-accuracy trade-off," where security comes at the cost of performance on clean data.

This research proposes a shift from purely mathematical defenses to those grounded in semantic reality. For instance, a "Stop" sign is defined by its octagonal shape. If an attack alters an image so much that its underlying geometry is no longer octagonal, it ceases to be a Stop sign in any meaningful physical sense. By enforcing these physical-world "invariances," we can significantly increase the effort required for an attacker to succeed.

The Hierarchical Classification Framework

The authors introduce a hierarchical paradigm that decomposes the classification task into manageable, robust steps:

  1. Invariance Extraction: Identifying salient features that must remain constant for a label to be valid (e.g., shape for traffic signs or gender for speaker ID).
  2. Intermediate Classifiers: These "root" nodes are trained for extreme robustness, classifying inputs into broad "equivalence classes" based on the identified invariances.
  3. Leaf Classifiers: Once the broad category is locked in, a specialized classifier makes the final fine-grained prediction.
  4. Robustness Certification: The framework utilizes Randomized Smoothing to provide a mathematical guarantee (a "certificate") of the radius within which the model's prediction remains stable.

Key Insights: The Margin Effect and Multi-Modality

A critical discovery is the "Margin Effect." Robustness is mathematically proportional to the gap between the most likely class and the runner-up. By using invariances to eliminate impossible classes (e.g., ignoring circular sign labels if an octagon is detected), the model naturally widens this margin.

Furthermore, the framework supports multi-modal defense. In driving scenarios, the authors combined camera data with LiDAR point clouds. Because LiDAR measures physical depth and geometry, it is inherently immune to 2D pixel-space attacks like stickers or lighting changes, providing a "ground truth" for shape detection that 2D-only models lack.

Experimental Results

The framework was validated across several domains:

  • Traffic Sign Recognition (GTSRB): The hierarchical approach consistently outperformed baselines. For circular and triangular signs, the certified robustness radius improved by over 80%.
  • Autonomous Driving (KITTI): By manually annotating thousands of images with LiDAR data, the authors demonstrated that physical constraints significantly harden models against real-world perturbations.
  • Speaker Identification (LibriSpeech): Using gender as an invariant, the model achieved a ~2x increase in certified robustness for both male and female speakers.

Crucially, this method often breaks the traditional trade-off, increasing both natural accuracy and adversarial robustness simultaneously. For example, on the CIFAR-10 dataset, the authors observed a 6.7% increase in adversarial accuracy over standard baselines.

Limitations and Future Outlook

While promising, the approach currently requires significant domain expertise to identify and label invariances manually. Future research is needed to automate the discovery of these constraints and to determine the optimal ordering of hierarchical nodes. Additionally, the subjective nature of human-interpretable invariances raises important questions regarding potential biases or fairness issues that must be addressed as these systems move toward real-world deployment.

Summary updated

BibTeX
@misc{chandrasekaran2019enhancing,
  author = {Varun Chandrasekaran and Brian Tang and Varsha Pendyala and Kassem Fawaz and Somesh Jha and Xi Wu},
  title = {{Enhancing ML Robustness Using Physical-World Constraints}},
  year = {2019},
  eprint = {1905.10900},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1905.10900},
}
Figure