Abstract
Deep Neural Networks (DNNs) are highly vulnerable to adversarial examples—inconspicuous perturbations that cause system failure. While gradient-based attacks are effective, they suffer from high latency and memory overhead. Generative-based attacks offer speed but often struggle with training stability and visual quality due to the complexities of Generative Adversarial Networks (GANs). This paper introduces the Symmetric Saliency-based Auto-Encoder (SSAE), a discriminator-free generative attack. By employing a saliency map module to target label-relevant regions and an angle-norm disentanglement module for feature manipulation, SSAE generates high-quality adversarial examples that effectively compromise various models across multiple tasks.
Introduction
Adversarial attacks pose a significant security risk to DNN-based systems. Traditionally, generating these attacks involves a trade-off between speed and quality. Gradient-based methods (e.g., PGD) are slow and computationally intensive. Conversely, generative methods like AdvGAN are fast but difficult to train, often resulting in poor convergence and visible "circular noise" artifacts in the generated images.
Researchers from Sun Yat-sen University and Tencent Youtu Lab propose the SSAE to overcome these hurdles. By removing the discriminator, they achieve a more stable training process while maintaining the inference speed of generative models.
Methodology
The core of the proposed approach is the Symmetric Saliency-based Auto-Encoder (SSAE). Unlike GANs, which rely on a discriminator to enforce realism, SSAE uses architectural constraints and feature-space manipulation to ensure both attack efficacy and visual stealth.
Symmetric Saliency-based Architecture
The SSAE architecture consists of a single encoder and two symmetric decoders:
- Noise Decoder ($G_N$): Generates the raw adversarial perturbation.
- Saliency Map Decoder ($G_M$): Learns an unsupervised mask that identifies regions most critical to the target model's classification decision.
The final adversarial perturbation is the element-wise product of the noise and the saliency map. This ensures that the noise is concentrated on the object itself, making it less noticeable to human observers while remaining highly disruptive to the model.
Angle-Norm Disentanglement
To improve the transferability of the attack, the authors introduce angle-norm disentanglement. They decompose the feature vector of an image into its angle and its norm (magnitude). The loss function is designed to push the angle of the perturbed image's feature vector away from the original while keeping the norm constant. This targets the fundamental orientation of features that models use to identify objects, allowing the attack to transfer effectively between different architectures.
Experimental Evaluation
The researchers evaluated SSAE across tasks including image classification (CIFAR-10, Imagenette) and Person Re-identification (Market1501, CUHK03).
- Visual Quality: SSAE produces significantly cleaner images than AdvGAN, achieving higher Structural Similarity (SSIM) and Peak Signal-to-Noise Ratio (PSNR) scores.
- Attack Strength: On the Imagenette dataset, SSAE reduced ResNet18 accuracy from 94.0% to 11.1%, whereas AdvGAN failed to significantly impact the same model.
- Efficiency: SSAE is 25x to 60x faster than iterative gradient-based methods like PGD during inference and consumes significantly less GPU memory.
- Transferability: An SSAE trained on ResNet18 successfully reduced the accuracy of a MobileNetv2 model by 76%, demonstrating robust cross-model performance.
Conclusion
The SSAE demonstrates that a discriminator is not necessary for effective generative adversarial attacks. By focusing on saliency-guided noise placement and feature-space angle manipulation, the method provides a fast, stable, and visually stealthy way to probe the vulnerabilities of Deep Neural Networks. Future work will explore refining saliency maps for complex backgrounds and investigating defensive measures against such feature-space attacks.
Paper Citation: Shaohao Lu et al.
- Discriminator-Free Generative Adversarial Attack. In Proceedings of the 29th ACM International Conference on Multimedia (MM ’21). https://doi.org/10.1145/3474085.3475290
Summary updated