Introduction
In the field of AI security, transfer-based black-box attacks represent a significant vulnerability. These attacks involve crafting adversarial examples (AEs) on a local surrogate model with the intent that they will "transfer" to successfully deceive an unknown target model. However, the effectiveness of these attacks is often limited by AE overfitting, where the attack exploits idiosyncratic "blind spots" or sharp local optima in the surrogate's loss landscape that do not generalize to other architectures.
The Lipschitz Regularized Surrogate (LRS)
To address this, we propose the Lipschitz Regularized Surrogate (LRS) framework. Unlike traditional methods that focus on refining the attack algorithm, LRS transforms the surrogate model itself to create a more favorable "landscape" for adversarial generation. By fine-tuning a pre-trained model for a few epochs with specific regularization, we can fundamentally alter its internal geometry.
LRS introduces three levels of regularization:
- LRS-1 (First-Order): Constrains the local Lipschitz constant by penalizing the $L_2$ norm of the loss gradient with respect to the input, preventing sharp spikes in the loss.
- LRS-2 (Second-Order): Penalizes the curvature (Hessian) of the landscape. To maintain computational efficiency, we employ a Finite Difference Method (FDM) approximation.
- LRS-F (Full): Combines both first- and second-order regularization for maximum smoothing.
Key Insights: Smoothing for Generalization
The core philosophy behind LRS is that a smoother loss landscape promotes "flat optima"—broad regions of high loss that are more likely to be shared across different neural network architectures. By reducing the local Lipschitz constant, the attack is forced to find perturbations that are robust across a wider neighborhood, enhancing their transferability. Empirical measurements confirm that LRS-transformed models are over an order of magnitude smoother than original pre-trained models.
Experimental Evaluation
We evaluated LRS on standard benchmarks including CIFAR-10 and ImageNet. The results demonstrate that LRS is highly compatible with existing state-of-the-art attack algorithms like PGD, TIM, and Admix.
On ImageNet, using a ResNet-50 surrogate, LRS-F achieved a remarkable increase in average Attack Success Rate (ASR), jumping from 27.69% to 69.91%. As shown in the table below, LRS consistently outperforms other established transferability-enhancing methods across various target architectures, including SENet, ResNeXt, and MNASNet.
Conclusion
LRS provides a powerful, model-agnostic "cushion" for adversarial attacks. By smoothing the surrogate's loss landscape through Lipschitz regularization, it significantly mitigates the problem of AE overfitting. This research highlights that the geometry of the surrogate model is just as critical as the attack algorithm itself in determining the success of black-box adversarial transfers.
Summary updated