Abstract
Deep Neural Networks (DNNs) are highly susceptible to adversarial attacks, where small, human-imperceptible perturbations lead to significant model failures. This paper proposes a novel "plug-and-play" defense mechanism that utilizes Deep Generative Networks (DGNs) to bridge the distribution gap between clean and adversarial images. By aligning these distributions at both the pixel and feature levels, the framework provides a robust, task-agnostic defense. Experimental results across image classification, semantic segmentation, and object detection demonstrate that this approach significantly outperforms existing state-of-the-art input transformation defenses.
Introduction
The vulnerability of DNNs to adversarial samples is a major hurdle for their deployment in security-critical applications. This research focuses on black-box attacks, which exploit the transferability of adversarial samples across different model architectures. The core challenge identified is a distribution gap: in a model's high-dimensional feature space, adversarial samples often fall into "disordered" regions far from the clean data distribution. While traditional defenses like denoising or adversarial training exist, they often struggle with optimization complexity or fail to generalize across diverse computer vision tasks.
Methodology
The proposed framework introduces a Deep Generative Network (DGN) that serves as a pre-processor to "clean" images before they are fed into a target model. The defense relies on two primary alignment strategies:
1. Pixel-Level Alignment
Rather than attempting the "hard" optimization of mapping an adversarial image $x^a$ directly back to its clean counterpart $x^c$, the authors map both to an intermediate "cleaned" space ($\hat{x}^a$ and $\hat{x}^c$). By minimizing the distance between these two generated versions, the DGN learns a smoother optimization landscape, facilitating faster convergence of the Proxy-A distance.
2. Class-Aware Feature Alignment
To ensure the cleaned images are useful for the downstream task, the DGN is guided by the target model's feature space using three specific losses:
- Alignment Loss: Pulls the centers of adversarial and clean feature clusters together.
- Intra-class Loss: Minimizes the variance within features of the same class.
- Inter-class Loss: Maximizes the separation between different classes to maintain clear decision boundaries.
The DGN is further refined using a Task-oriented loss ($\mathcal{L}_m$), ensuring that the generated images preserve the semantic features necessary for specific tasks like object detection or segmentation.
Distribution Alignment and Visual Insights
The fundamental insight of this work is that effective defense requires matching the entire distribution of adversarial samples to the clean distribution, rather than just correcting individual points.
As shown in the t-SNE visualizations above, adversarial samples (b) exhibit a highly disordered distribution compared to clean samples (a). The proposed defense successfully restores the clustered structure in the "cleaned" adversarial space (d), making it nearly indistinguishable from the cleaned clean space (c).
Experimental Results
The authors validated their approach on standard benchmarks including CIFAR10/100, Tiny-ImageNet, Cityscapes, and Pascal VOC.
- Image Classification: On CIFAR10 under PGD attack, the method achieved 81.4% accuracy, surpassing the previous state-of-the-art (NRP) at 79.0%.
- Semantic Segmentation: On the Cityscapes dataset, where an undefended model's performance dropped to 3.6% mIoU under attack, this method maintained a robust 59.5% mIoU.
- Object Detection: On VOC07+12, the framework achieved 61.2% mAP against classification-based attacks, a massive improvement over the 17.9% baseline.
A key finding is the defense transferability: a DGN trained to protect one specific model (e.g., WideResNet) can effectively shield a completely different, unseen model (e.g., ResNet50), making it a practical solution for real-world "plug-and-play" deployment.
Conclusion
This paper presents a powerful distribution-alignment strategy for adversarial defense. By addressing the distribution gap at both pixel and feature levels, the DGN-based approach offers a task-agnostic solution that generalizes well across various computer vision domains. While future work may explore white-box scenarios and computational latency for real-time applications, the current results establish a new benchmark for robust black-box defense.
Summary updated