Gadt: Enhancing Transferable Adversarial Attacks Through Gradientguided Adversarial Data Transformation

Yating Ma, Xiaogang Xu, Liming Fang, Jiafei Wu, Lu ZhouSSRN 2025 · DOI

TL;DR

This paper introduces GADT, a novel method for generating more effective transferable adversarial examples (TAEs) by optimizing data augmentation (DA) parameters using gradient information. By treating DA operations as differentiable components, GADT can fine-tune them to maximize attack potency while maintaining image stealth, significantly boosting the success rate of existing attack methods.

What Problem Does This Paper Address?

Deep neural networks (DNNs) are increasingly used in critical applications, but they remain vulnerable to adversarial examples (AEs) – subtly modified inputs that cause misclassifications. Transferable attacks (TAs) are particularly concerning because they require minimal knowledge of the target model, making them practical for real-world black-box scenarios. While existing TA methods primarily focus on generating adversarial noise (AN), recent research highlights the importance of data augmentation (DA) in enhancing attack effectiveness. However, optimizing DA parameters alongside AN is challenging due to the complex search space and lack of direct gradient guidance. This paper addresses this gap by proposing a gradient-guided approach to optimize DA parameters, making TAs more potent and stealthy.

Methods

GADT introduces a two-stage attack pipeline:

  1. Data Augmentation Parameter Optimization via Gradients: This stage focuses on optimizing the parameters of differentiable DA operations. Instead of relying on empirical tuning or extensive search, GADT directly computes the gradient of an attack-specific loss function with respect to the DA parameters. This allows for iterative updates to these parameters, guiding them towards configurations that maximize adversarial effects. The paper utilizes Kornia, a differentiable computer vision library, to implement operations like motion blur and saturation adjustment with adjustable parameters.
  2. Adversarial Noise Formulation based on Optimized DA Parameters: Once the optimal DA parameters are identified, the transformed images are fed into existing AN-based attack methods (e.g., MI-FGSM). This ensures that the generated adversarial examples benefit from both the enhanced data diversity from optimized DA and the perturbation strategies of established AN-based attacks.

A key innovation is the novel loss function, $L_{trans} = -L_{CE}(f(x_{trans}), y) + \lambda \cdot L_{MSE}(x, x_{trans})$, which balances attack efficacy (via Cross-Entropy loss, $L_{CE}$) with image fidelity (via Mean Squared Error loss, $L_{MSE}$). This dual objective helps maintain the stealthiness of adversarial examples.

Datasets

The experiments were conducted on a dataset of 1,000 images extracted from an ImageNet-compatible dataset, commonly used in adversarial competitions.

Key Insights and Ideas

  • Gradient-guided DA optimization is crucial: Directly optimizing DA parameters using gradients, rather than relying on heuristic search or fixed transformations, significantly enhances attack effectiveness.
  • Differentiable DA operations are enabling: The availability of libraries like Kornia, which provide differentiable DA operations, is key to enabling gradient-based optimization of augmentation parameters.
  • Balancing potency and stealth is achievable: The proposed loss function effectively balances the need for strong adversarial effects with the requirement for imperceptible perturbations, leading to more robust and stealthy attacks.
  • GADT is a versatile framework: The method is designed to be integrated with various existing transferable attack methods, acting as a plug-in to improve their performance.

Key Results

  • GADT consistently improves the performance of baseline attack methods (MIM, SIM, Admix) across multiple surrogate and target models, often increasing attack success rates by 5% to 20%.
  • When combined with TIM, GADT achieved significant improvements against defended models (Inc-v3adv, AT, HGD), outperforming existing SOTA methods.
  • GADT demonstrated strong transferability to vision-language models like CLIP, showing improvements of over 10%.
  • The proposed loss function ($L_{trans}$) was shown to maintain high fidelity between adversarial and clean samples, as measured by PSNR and SSIM metrics.

Limitations and Future Work

  • Limited DA operations: The current implementation of GADT uses only two differentiable DA operations (motion blur and saturation adjustment) due to the limitations of available libraries. Future work could explore a wider range of differentiable DA operations.
  • Computational cost: The iterative optimization of DA parameters adds to the overall attack generation time, although the paper suggests it is more efficient than extensive search-based methods.
  • Generalizability to other black-box attacks: While the paper mentions potential for other black-box attacks (e.g., query-based), further exploration and validation in those domains would be beneficial.
  • Defense mechanisms: While GADT shows promise against some defenses, further research into its effectiveness against more sophisticated adversarial defense strategies is warranted.

Summary updated

BibTeX
@misc{ma2025gadt,
  author = {Yating Ma and Xiaogang Xu and Liming Fang and Jiafei Wu and Lu Zhou},
  title = {{Gadt: Enhancing Transferable Adversarial Attacks Through Gradientguided Adversarial Data Transformation}},
  year = {2025},
  doi = {10.2139/ssrn.5342331},
  howpublished = {SSRN},
}
Figure