Abstract
Transferability of adversarial examples is of critical importance to launch black-box adversarial attacks, where attackers are only allowed to access the output of the target model. However, under such a challenging but practical setting, crafted adversarial examples are often prone to overfitting the proxy model employed, resulting in poor transferability. This paper proposes alleviating the overfitting issue through a novel perspective: designing a fitted model architecture. Specifically, we decompose and reconstruct existing model architectures into an effective multi-track model architecture (MMA). Adversarial examples crafted on the MMA minimize the effect of model-specific features and target vulnerable directions shared by diverse architectures. Extensive evaluation demonstrates that MMA significantly surpasses other state-of-the-art architectures, improving transferability by up to 40% with comparable overhead.
Introduction
Adversarial attacks—where tiny, invisible perturbations to an image cause a machine learning model to misclassify it—pose a significant security concern. In a "black-box" scenario, the attacker lacks knowledge of the target model's architecture or weights. To circumvent this, attackers typically train a "proxy model," craft an attack against it (white-box), and hope the resulting adversarial example "transfers" to the target. The primary obstacle is overfitting: adversarial examples often become too specialized to the proxy model's unique high-level features, leading to low success rates on unseen target models.
Methods: The Multi-track Model Architecture (MMA)
The core insight of the Multi-track Model Architecture (MMA) is that transferability stems from shared low-level features (like edges and basic textures) that most models learn similarly, whereas high-level semantic features are often model-specific.
Key components of MMA include:
- Decomposition and Reconstruction: Instead of a single deep monolithic network, MMA splits the architecture into multiple parallel tracks or small networks.
- Multi-Head Output: The architecture features multiple "heads" at different depths, allowing the attacker to generate adversarial examples from various stages of the network.
- Feature Sharing: Tracks are not entirely independent; networks placed further back in the architecture can utilize feature maps from those in the front, balancing shared feature learning with model-specific noise reduction.
- Ensemble Gradient: By crafting attacks based on the ensemble output of these multiple heads, the resulting adversarial noise is guided toward vulnerable directions common across diverse architectures.
Key Results and Insights
The researchers evaluated MMA using the CIFAR-10 dataset against widely-used architectures including MobileNet, GoogleNet, and ResNet18. The MMA approach showed dominant performance across both single-step (FGSM) and multi-step (BIM) attacks.
Key findings from the experimental results include:
- Superior Transferability: MMA improved transferability by 2.32% to 17.96% in single-step attacks. In multi-step attacks, the improvements were more dramatic, achieving a maximum Attack Success Rate (ASR) of 79.79%—an improvement of up to 40% over standard architectures.
- The "Sweet Spot" of Depth: Transferability is not linear with depth. A depth of 3 blocks was found to be the "crest"; increasing depth beyond this point damaged transferability as the model began learning too many exclusive, high-level features.
- Training Dynamics: ASR typically peaks early (around epoch 10) and then declines. This suggests that as a model becomes more accurate and specialized, it becomes a worse proxy for transfer attacks.
- Efficiency: Despite its multi-track nature, MMA's computational overhead is comparable to standard baseline models, making it a practical choice for real-world attack scenarios.
Conclusion
This work demonstrates that the architecture of a proxy model itself can act as a form of regularization to prevent adversarial overfitting. By prioritizing shared features through a multi-track design, MMA significantly enhances the transferability of adversarial examples. Future work may explore scaling MMA to higher-resolution datasets like ImageNet and integrating it with other transfer-boosting techniques.
Summary updated