Abstract
Adversarial attacks pose a significant threat to the reliability of deep learning models. While adversarial training and purification methods offer defenses, they often suffer from performance trade-offs or high computational costs. This paper introduces the Feature Pattern Consistency Constraint (FPCC), a "clean training" approach that improves intrinsic robustness without using adversarial examples during training. FPCC leverages Spatial-wise Feature Modification (SFM) and Channel-wise Feature Selection (CFS) to simulate distortions, while Pattern-based Robustness Optimization (PRO) enforces consistency with a ground-truth feature pattern. Experimental results on CIFAR benchmarks demonstrate that FPCC achieves state-of-the-art robust accuracy while maintaining high standard accuracy and zero inference overhead.
Introduction
Deep learning models, particularly Convolutional Neural Networks (CNNs), are notoriously fragile. "Adversarial attacks"—tiny, human-imperceptible perturbations added to an image—can easily trick a model into misclassification. Current defenses usually fall into two categories: adversarial training, which often leads to overfitting on specific attacks and reduced clean accuracy, and adversarial purification, which introduces significant latency during inference.
The authors tackle the "clean training" challenge: building intrinsic robustness using only normal images. The goal is to develop a model that is robust without prior knowledge of the attack type and without sacrificing inference speed.
Methodology
The core philosophy of FPCC is that a correct prediction depends on the Feature Pattern—the relative values across different dimensions in a network's latent layers. By focusing on the structural essence of a category rather than absolute pixel intensities, the model becomes more resilient to manipulation.
The FPCC framework incorporates three main components during the training phase:
- Spatial-wise Feature Modification (SFM): This module adds random noise to the input features, simulating the "interferential features" introduced by adversarial attacks.
- Channel-wise Feature Selection (CFS): This module randomly drops certain feature channels to simulate the loss of critical signals that occurs when an attack suppresses the features the model typically relies on.
- Pattern-based Robustness Optimization (PRO): This is the core loss function. It calculates a "ground-truth" feature pattern for each category by averaging successful clean examples. It then uses an L1 distance constraint to force the model's features to match this ground-truth pattern, even when SFM and CFS are introducing distortions.
To focus on relative relationships, features are normalized using z-scores to create the pattern. Crucially, these modifications are only applied during training; at test time, the model functions as a standard, efficient CNN.
Key Results
The authors evaluated FPCC on the CIFAR-10 and CIFAR-100 benchmarks using popular architectures like WRN-28-10, ResNet-50, and VGG-16.
FPCC achieved impressive results, often outperforming other "clean training" methods and rivaling traditional adversarial training:
- Superior Robustness: On CIFAR-10 (WRN-28-10), FPCC achieved a robust accuracy of 65.58% against PGD attacks, significantly higher than other clean-training methods like RHS or MMC.
- Maintained Standard Accuracy: Unlike adversarial training, which often degrades performance on normal images, FPCC maintained a high clean accuracy of 88.64% on CIFAR-10.
- Efficiency: Because the defense is "baked in" to the model's weights, inference speed is identical to a standard, undefended model, making it much faster than purification-based methods.
- Genuine Robustness: Sanity checks on gradient norms confirmed that the robustness is intrinsic and not due to "gradient obfuscation," a common pitfall in defense research.
Conclusion
FPCC offers a powerful and efficient method for building adversarial robustness into neural networks without the need for adversarial examples during training. By enforcing feature pattern consistency, the model learns to identify the essential characteristics of a category that are difficult for attackers to disrupt. Future research will focus on providing a more rigorous theoretical foundation for this approach and scaling it to larger datasets like ImageNet.
Summary updated