Abstract
Deep Neural Networks (DNNs) are increasingly deployed on edge devices for tasks like semantic segmentation in autonomous driving. However, these models remain vulnerable to adversarial attacks. This study introduces a non-linear adversarial attack method designed to overcome gradient truncation and low transferability across different lightweight architectures. By employing a tanh mapping function and a joint-learning optimization strategy, the proposed method effectively compromises multiple models simultaneously while maintaining human-imperceptible perturbations.
Introduction
Semantic segmentation, the process of assigning a class label to every pixel in an image, is critical for the perception systems of autonomous vehicles. While adversarial attacks on image classification are well-documented, attacking segmentation is more complex due to the high dimensionality of the output space. Existing methods often struggle with gradient truncation—where pixel values are clipped at boundaries—and poor transferability between diverse architectures like ESPNet and FastSCNN.
Methodology
The researchers propose a novel framework that addresses these limitations through two primary innovations: non-linear perturbation generation and joint-learning optimization.
Non-linear Perturbation Generation
Traditional attacks like the Fast Gradient Sign Method (FGSM) use hard clipping to keep adversarial images within valid pixel ranges (e.g., [0, 255]). This clipping often results in zero gradients for pixels at the boundaries, hindering the optimization process. To solve this, the authors generate perturbations indirectly using a tanh function mapping.
By mapping intermediate variables through this non-linear function, the gradient can propagate without being cut off, allowing for more effective exploration of the adversarial space and more potent attacks on pixels at extreme brightness levels.
Joint Learning for Multi-Model Attacks
To improve transferability, the study employs a joint optimization strategy. Instead of targeting a single model, the attack minimizes the performance of multiple architectures simultaneously. A unified ground truth map is formulated based on the consensus of the targeted models, ensuring the generated noise exploits vulnerabilities common to different architectural designs, such as ESPNet's dilated convolutions and FastSCNN's multi-resolution fusion.
Experimental Results
The method was evaluated using the Cityscapes dataset, focusing on 500 urban images. The results demonstrate a significant improvement over baseline attacks:
- Accuracy Reduction: In joint attacks, target class accuracy dropped to near zero (e.g., 0.002 for ESPNet), compared to 0.133 for the standard FGSM.
- Success Rate: The pixel-flip success rate (the ratio of pixels successfully changed to the adversarial class) often exceeded 70% in complex scenarios.
- Stealthiness: Despite the high success rate, the perturbations remained extremely low and nearly imperceptible to the human eye, maintaining the "stealth" required for a successful adversarial attack.
Discussion and Conclusion
A key insight from the study is the "receptive field effect," where local perturbations influence neighboring regions due to the convolutional nature of the models. The researchers found that different lightweight models have distinct "blind spots"; for instance, ESPNet is more sensitive to global features while FastSCNN is sensitive to local details.
While highly effective in white-box settings where model gradients are known, future work is required to extend these techniques to black-box scenarios. The authors emphasize that these findings should be used to develop robust adversarial training defenses to protect lightweight models in real-world, safety-critical applications.
Summary updated