Looking From the Future: Multi-order Iterations Can Enhance Adversarial Attack Transferability

Zijian Ying, Qianmu Li, Tao Wang, Zhichao Lian, Shunmei Meng, Xuyun ZhangarXiv 2024 · DOI · arXiv

Abstract

Adversarial attacks often suffer from poor transferability, where perturbations crafted for a specific "white-box" model fail to deceive "black-box" targets. This is largely due to the "short-sighted" nature of current iterative attacks that rely solely on local or historical gradients, leading to overfitting on the source model's architecture. We introduce Looking From the Future (LFF), a novel optimization strategy that simulates subsequent steps to refine the current update. By incorporating "future" gradient information, LFF acts as a regularizer that helps the optimization process escape local minima. Our extensive evaluation on ImageNet-1k shows that LFF consistently boosts the transferability of state-of-the-art attacks across various CNN and Transformer architectures, achieving up to a 14.6% improvement in attack success rates against defensive models.

Introduction

The vulnerability of Deep Neural Networks (DNNs) to adversarial examples poses significant security risks. A critical challenge in adversarial research is enhancing transferability: the ability of an attack generated on a known model to remain effective against unknown black-box models. Existing iterative methods, such as I-FGSM and its momentum-based variants (MI-FGSM), typically determine the next perturbation step based only on the current state and past gradients. This "greedy" approach often traps the optimization in narrow local minima unique to the white-box model, resulting in poor generalization.

Methodology: Looking From the Future

The core innovation of this work is the "look-ahead" mechanism. Instead of a simple gradient update, the LFF strategy simulates the optimization path several steps into the future.

LFF Variants

  1. One-Order LFF Attack: The algorithm simulates $Q$ future steps and updates the current perturbation using a weighted combination of these future gradients.
  2. Momentum LFF (MLFF): This variant integrates LFF with momentum-based attacks. The "future-aware" gradients are used to update the momentum buffer, effectively combining historical trends with future projections.
  3. Multi-order LFF ($LFF^N$): This treats the LFF process itself as an iterative sequence, allowing for even deeper refinement of the perturbation path.

A critical component is the future penalty coefficient ($\beta$), which weights the influence of future gradients. This ensures that the attack doesn't drift too far from the local optimization goal while still benefiting from the look-ahead regularization.

Experimental Setup

We evaluated the proposed methods using the ImageNet-1k dataset. The study included a diverse set of victim architectures:

  • CNNs: ResNet-50, BiT-50, Inception-v3, Inception-ResNet v2, and ConvNeXT-B.
  • Transformers: ViT-B, Swin-B, and DeiT-B.
  • Defensive Models: Adversarially trained models (e.g., Inc-v3_adv) and those employing feature denoising (FD) or bit-reduction.

Results and Discussion

LFF-enhanced attacks consistently outperformed baseline state-of-the-art (SOTA) methods. As shown in the comprehensive results table below, the MLFF variant provides a significant boost to various base attacks like MI, EMI, Admix, and SIA.

Table 1: Transferability of MLFF-enhanced attacks across various architectures.

Key Findings

  • Significant ASR Gains: When applied to the Scale-Invariant Attack (SIA), the MLFF-SIA variant increased the Attack Success Rate (ASR) against black-box models by up to 12%. Against defensive models, the improvement reached 14.6% over the baseline.
  • Cross-Architecture Robustness: The method proved effective regardless of whether the source model was a CNN (Inception-v3) or a Transformer (ViT-B).
  • Ensemble Effectiveness: In ensemble settings, MLFF-SIA achieved near-perfect (99%+) transferability across multiple victim architectures.

The Role of Hyperparameters

The effectiveness of LFF is sensitive to the choice of $\beta$ and the number of look-ahead steps $Q$. The authors found an "optimal peak" for $\beta$ around 0.9 to 1.1.

Figure 1: Impact of the future penalty coefficient beta and look-ahead steps Q on Attack Success Rate (ASR).

As illustrated in Figure 1, if $\beta$ is too small, the benefit of looking ahead is lost. Conversely, a $\beta$ that is too large causes the attack to over-generalize, missing the specific target features of the source model. The parameter $Q$ also plays a role; while larger $Q$ values generally improve performance, they also increase computational cost.

Limitations and Future Work

The primary limitation of LFF is its computational overhead. Simulating $Q$ steps into the future requires $Q$ additional gradient calculations per iteration, making the attack slower than standard methods. Additionally, the benefit of LFF is somewhat reduced when combined with highly stochastic methods like Admix, as the randomness makes the "future" path less predictable. Future research could focus on more efficient ways to approximate future gradients or adaptive weighting schemes for $\beta$.

Conclusion

Looking From the Future (LFF) represents a shift from greedy optimization to a more regularized, forward-looking approach in adversarial attacks. By mitigating overfitting to the white-box source model, LFF significantly enhances the transferability of perturbations, posing a more robust challenge for black-box defense evaluation.

Summary updated

Abstract

Various methods try to enhance adversarial transferability by improving the generalization from different perspectives. In this paper, we rethink the optimization process and propose a novel sequence optimization concept, which is named Looking From the Future (LFF). LFF makes use of the original optimization process to refine the very first local optimization choice. Adapting the LFF concept to the adversarial attack task, we further propose an LFF attack as well as an MLFF attack with better generalization ability. Furthermore, guiding with the LFF concept, we propose an $LLF^{\mathcal{N}}$ attack which entends the LFF attack to a multi-order attack, further enhancing the transfer attack ability. All our proposed methods can be directly applied to the iteration-based attack methods. We evaluate our proposed method on the ImageNet1k dataset by applying several SOTA adversarial attack methods under four kinds of tasks. Experimental results show that our proposed method can greatly enhance the attack transferability. Ablation experiments are also applied to verify the effectiveness of each component. The source code will be released after this paper is accepted.
BibTeX
@misc{ying2024looking,
  author = {Zijian Ying and Qianmu Li and Tao Wang and Zhichao Lian and Shunmei Meng and Xuyun Zhang},
  title = {{Looking From the Future: Multi-order Iterations Can Enhance Adversarial Attack Transferability}},
  year = {2024},
  eprint = {2407.01925},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2407.01925},
}
Figure