D-BADGE: Decision-Based Adversarial Batch Attack With Directional Gradient Estimation

Geunhyeok Yu, Minwoo Jeon, Hyoseok Hwang

Abstract

Deep Neural Networks (DNNs) are susceptible to Universal Adversarial Perturbations (UAPs)—single noise patterns that can cause misclassification across an entire dataset. However, crafting UAPs in decision-based black-box settings, where only the final predicted label is accessible, remains a significant challenge due to the discrete nature of the feedback. This paper presents Decision-BADGE, a novel framework that optimizes UAPs by aggregating binary decisions into a continuous "Batch Accuracy Loss." By utilizing a refined stochastic gradient estimation technique, SPSA with Adaptive Momentum (SPSA-AM), Decision-BADGE achieves state-of-the-art attack success rates and superior computational efficiency compared to existing methods.

1. Introduction

The vulnerability of Deep Neural Networks to adversarial examples is well-documented. Most existing research focuses on white-box settings, where the attacker has full access to model parameters, or score-based black-box settings, where class probabilities are available. In contrast, real-world AI services often operate under a decision-based black-box constraint, returning only a discrete label. This lack of continuous feedback makes gradient-based optimization difficult. Furthermore, most decision-based attacks are image-specific, requiring a new optimization process for every target image.

Decision-BADGE addresses these limitations by focusing on Universal Adversarial Perturbations (UAPs) in the decision-based setting. The goal is to find a single perturbation that, when added to any input, causes the model to fail, using only the model's final classification decisions as feedback.

2. Methodology

The core of Decision-BADGE is the transformation of discrete "Yes/No" labels into a continuous optimization signal through two primary innovations.

2.1 Batch Accuracy Loss

Rather than evaluating the success of a perturbation on a single image, Decision-BADGE considers a mini-batch of data. The authors propose a "Batch Accuracy Loss" based on the Hamming distance between the model's predicted labels and the ground truth distribution. This aggregation provides a richer, more informative signal than a single binary output, effectively turning a discrete problem into a continuous, differentiable-like landscape. The authors mathematically demonstrate that this loss function is both convex and 1-Lipschitz-continuous, providing a solid theoretical foundation for optimization via stochastic approximation.

2.2 SPSA with Adaptive Momentum (SPSA-AM)

To estimate gradients without direct access to the model's internals, the framework employs Simultaneous Perturbation Stochastic Approximation (SPSA). To overcome the inherent noise and slow convergence of standard SPSA, the authors integrate adaptive momentum (Adam optimization). This SPSA-AM approach allows the algorithm to retain information from successful update directions, leading to more stable convergence and a significant reduction in the number of queries required to find an optimal perturbation.

3. Experimental Evaluation

The researchers evaluated Decision-BADGE against several victim architectures, including ResNet18/20, VGG19, MobileNet v2, ResNeXt29, and Vision Transformers (ViT-T, Swin-T). Testing was conducted on CIFAR-10, CIFAR-100, and MNIST datasets.

3.1 Performance and Efficiency

Decision-BADGE demonstrated superior performance across benchmarks. On the ResNet18 architecture, the method achieved an Attack Success Rate (ASR) of 85.88%, surpassing the previous state-of-the-art decision-based attack, DUAttack (85.10%), and even the white-box DF-UAP method (73.99%).

Efficiency is a standout feature of the framework. Decision-BADGE reached its peak ASR on ResNet18 in approximately 6,541 seconds, whereas DUAttack required 28,852 seconds. This represents a nearly 4.5x speedup in the time required to generate an effective universal perturbation.

3.2 Targeted Attacks and Transferability

The framework is also effective for targeted attacks, where the goal is to force the model into a specific incorrect class (e.g., misclassifying various animals as "planes"). The authors reported average target accuracies between 70% and 88%.

In terms of transferability, the study revealed a distinct "architecture gap." UAPs optimized for CNNs (like ResNet) transferred well to other CNNs but were less effective against Vision Transformers (ViTs). Conversely, UAPs trained on ViTs struggled against CNNs. Interestingly, the Swin Transformer exhibited partial transferability to both CNN and ViT architectures, suggesting it may share structural vulnerabilities with both paradigms.

4. Conclusion and Future Work

Decision-BADGE proves that universal adversarial attacks are highly feasible even in the most restrictive black-box settings. By leveraging batch-level feedback and adaptive gradient estimation, the method achieves high success rates with unprecedented efficiency.

Despite these gains, challenges remain. The "training" phase of the UAP still requires a high volume of queries to the victim model. Future research will focus on further reducing this query intensity. Additionally, maintaining high attack success rates while reducing the visibility of perturbations (lowering the $l_\infty$ norm) and improving cross-domain transferability between CNNs and Transformers are key areas for ongoing development.

Summary updated

BibTeX
@article{yu2024d,
  author = {Geunhyeok Yu and Minwoo Jeon and Hyoseok Hwang},
  title = {{D-BADGE: Decision-Based Adversarial Batch Attack With Directional Gradient Estimation}},
  journal = {IEEE Access},
  year = {2024},
  doi = {10.1109/access.2024.3407097},
}

Loading related content...

Figure