Towards Better Accuracy and Robustness with Localized Adversarial Training

Eitan Rothberg, Tingting Chen, Hao JiAAAI 2019 · DOI

Introduction

Deep learning models, particularly Convolutional Neural Networks (CNNs), face a significant security challenge: adversarial attacks. These small, human-imperceptible perturbations can cause models to misclassify images with high confidence. While Adversarial Training—training on malicious examples—is a standard defense, it traditionally introduces a "no free lunch" trade-off. As models become more robust, their accuracy on clean, natural images typically declines. This paper investigates Localized Adversarial Training (LAT) as a method to break this trade-off by focusing perturbations on non-essential image features.

Methodology: Localized Adversarial Training

The core hypothesis of this research is that standard adversarial training fails because it perturbs the entire image, including features essential for classification. This causes the model to lose focus on the primary object. LAT addresses this by restricting perturbations to specific areas, primarily the background.

Visual representation of the masking strategies (top), original images (middle), and the resulting adversarial examples (bottom) used during training.

The process involves generating noise via Projected Gradient Descent (PGD) and applying an epsilon matrix (epmatrix) mask to localize the noise. The researchers evaluated four masking strategies on the MNIST dataset:

  • Standard PGD: No mask (global perturbation).
  • General Background: Leaves the center 36 pixels untouched.
  • Exact Mask: Perturbs only the non-digit pixels.
  • Broad Mask: Leaves the digit and its immediate neighbors untouched.

Key Insights and Results

Experimental results reveal that CNNs are dangerously sensitive to background information. An "Exact Mask" attack, which leaves the digit perfectly intact but alters the background, fooled a standard natural model 99.92% of the time. By utilizing LAT, models are forced to ignore peripheral noise and rely on the invariant features of the actual object.

Comparative performance table showing scaled loss across different models and attack types. Note that the Broad Mask model achieves the lowest loss on natural images.

As shown in the results table, the Broad Mask model successfully broke the accuracy-robustness trade-off:

  • Improved Natural Accuracy: It achieved a scaled loss of 38.7 on natural images, outperforming the standard natural classifier (41.4) by approximately 6.5%.
  • Enhanced Robustness: It maintained at least 65.3% accuracy across all adversarial attacks, whereas the standard model's accuracy plummeted under similar conditions.
  • Efficiency: Unlike standard adversarial training, which increased natural image loss to 57.9, the Broad Mask approach served as a form of data augmentation that clarified the boundary between object and noise.

Conclusion and Future Work

Localized Adversarial Training demonstrates that it is possible to improve both robustness and accuracy simultaneously by guiding the model's attention toward core features. While these results on MNIST provide a strong proof of concept, future work must address the challenges of more complex datasets like CIFAR-10 or ImageNet. Implementing LAT on colorful, complex images will require integrating object detection or segmentation to generate effective masks, potentially making LAT a standard technique for secure and accurate AI.

Summary updated

BibTeX
@inproceedings{rothberg2019towards,
  author = {Eitan Rothberg and Tingting Chen and Hao Ji},
  title = {{Towards Better Accuracy and Robustness with Localized Adversarial Training}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.1609/aaai.v33i01.330110017},
}
Figure