DAASH: A Meta-Attack Framework for Synthesizing Effective and Stealthy Adversarial Examples

Abdullah Al Nomaan Nafi, Habibur Rahaman, Zafaryab Haider, Tanzim Mahfuz, Fnu Suya, Swarup Bhunia, Prabuddha ChakrabortyarXiv 2025 · DOI · arXiv

Abstract

Adversarial attacks often face a trade-off between mathematical effectiveness and visual stealth. Traditional $\ell_p$-norm constraints are poor proxies for human perception, leading to noticeable artifacts. We introduce DASH (Differential Attack SearcH), a meta-attack framework that learns to compose multiple base attacks into an optimized multi-stage pipeline. By utilizing a differentiable soft attention mechanism and a perceptual loss function, DASH achieves near-100% success rates against robust models while maintaining high visual quality. Evaluations on CIFAR and ImageNet show that DASH outperforms existing perceptual attacks like AdvAD and DiffAttack in both success rate and image fidelity.

Introduction

The security of neural networks is challenged by adversarial examples—inputs specifically crafted to cause misclassification. While early attacks relied on simple $\ell_p$-norm constraints to ensure small perturbations, these metrics often fail to align with human visual perception. A perturbation that is small in $\ell_p$ distance can still be highly visible if it disrupts smooth image regions. Conversely, larger changes in textured areas might remain unnoticed. DASH addresses this by bridging the gap between mathematical robustness and perceptual stealth, creating a "meta-attack" that manages a suite of existing methods to produce superior adversarial examples.

Methodology

DASH operates as a manager for a pool of base attacks, such as PGD, CW, and FGSM. Its architecture is defined by three core components:

  1. Multi-Stage Chaining: The attack is structured as a sequence of "attack cells." This iterative refinement allows the optimization process to escape local minima that often trap single-stage attacks.
  2. Differentiable Soft Attention: In each stage, DASH generates candidates from all available base attacks and uses a learned attention mechanism to blend them. This allows the framework to dynamically select the most effective attack strategy for different stages of the perturbation process.
  3. Meta-Loss Function: The system optimizes a combined objective that maximizes attack success (reducing model confidence in the correct class) while minimizing perceptual distortion using the Structural Similarity Index (SSIM).

Experimental Evaluation

The framework was rigorously tested on CIFAR-10, CIFAR-100, and ImageNet-1k datasets. To assess its performance against modern defenses, DASH was evaluated against seven state-of-the-art robust models from RobustBench and several post-processing defenses, including JPEG compression and denoising.

Results and Discussion

DASH demonstrates a significant performance leap over existing perceptual attacks. On robust CIFAR-100 models, DASH achieved an average Attack Success Rate (ASR) of 99.77%, vastly exceeding AdvAD’s 79.14%.

Visual comparison of DASH against state-of-the-art perceptual attacks. DASH maintains high visual fidelity while achieving superior attack success.

Beyond raw success rates, DASH produces images with superior visual quality. It achieved higher SSIM scores (94.43 vs. 83.18 for AdvAD) and lower LPIPS distances. Analysis of the attention weights reveals that DASH strategically shifts its reliance between different base attacks; for instance, it often starts with the Carlini-Wagner (CW) attack for initial misclassification before shifting to other methods to refine perceptual quality. Furthermore, the multi-stage approach proved critical, with ASR jumping from ~52% to ~99% when moving from one to two stages.

Conclusion

DASH represents a significant advancement in adversarial research by demonstrating that chaining and learning to compose existing attacks is more effective than developing isolated methods. While the computational cost is higher due to the multi-stage nature, the resulting attacks are both more potent and harder to detect. Future work will focus on optimizing hyperparameters within the base attacks and extending the framework to non-classification tasks such as object detection and multimodal analysis.

Summary updated

BibTeX
@misc{nafi2025daash,
  author = {Abdullah Al Nomaan Nafi and Habibur Rahaman and Zafaryab Haider and Tanzim Mahfuz and Fnu Suya and Swarup Bhunia and Prabuddha Chakraborty},
  title = {{DAASH: A Meta-Attack Framework for Synthesizing Effective and Stealthy Adversarial Examples}},
  year = {2025},
  eprint = {2508.13309},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.13309},
}
Figure