Efficient Ensemble Adversarial Attack for a Deep Neural Network (DNN)-Based Unmanned Aerial Vehicle (UAV) Vision System

Zhun Zhang, Qihe Liu, Shijie Zhou, Wenqi Deng, Zhewei Wu, Shilin QiuDrones 2024 · DOI

TL;DR

This paper introduces a novel black-box adversarial attack method for UAV vision systems that achieves high success rates with an exceptionally low number of queries (1-2 per image). By decomposing images into frequency bands and strategically applying perturbations, the method effectively deceives deep neural networks, posing a significant threat to their reliability.

What Problem Does This Paper Address?

Unmanned Aerial Vehicles (UAVs) increasingly rely on Deep Neural Networks (DNNs) for their vision systems to perform tasks like object recognition and navigation. While these DNNs offer impressive performance, they are also vulnerable to adversarial attacks. These attacks involve subtly altering input images to cause DNNs to misclassify them, which can have critical consequences for UAV operations. Traditional "black-box" adversarial attacks (where the attacker doesn't know the internal workings of the target DNN) often require a large number of queries to the target model to generate effective adversarial examples. This high query count makes them detectable and impractical in real-world scenarios. This paper aims to bridge this gap by developing an efficient black-box adversarial attack that achieves high success rates with minimal interaction with the target UAV vision system.

Methods

The core of the proposed method is a Frequency-Based Perturbation Generator (FPG). This generator works in several key steps:

  1. Frequency Decomposition: The input image is decomposed into four frequency bands (approximation, horizontal, vertical, and diagonal) using Discrete Wavelet Transform (DWT). This allows for targeted perturbations in different aspects of the image's information.
  2. Weighted Ensemble Loss Optimization: The FPG utilizes a set of pre-trained "surrogate" DNN models. It minimizes a weighted ensemble loss function across these surrogate models to guide the generation of perturbations that are likely to deceive a variety of models.
  3. Frequency Band Weights Optimization: An adaptive optimization approach, inspired by the Adam algorithm, is used to adjust the weights assigned to each frequency band. This allows the attack to focus on the bands that are most effective for deceiving the target model.
  4. Perturbation Addition: Perturbations are added to the selected frequency bands. This includes both gradient-based perturbations (guided by the surrogate models) and random block perturbations in high-gradient regions to enhance attack effectiveness.
  5. Reconstruction: The perturbed frequency bands are reconstructed back into an adversarial image using Inverse Discrete Wavelet Transform (IDWT).

The overall attack strategy is Frequency-Based Surrogate Ensemble Black-Box Attack (FSEBA). It combines query-based and transfer-based approaches. The FPG generates adversarial samples using surrogate models, and these samples are then queried against the target UAV model. The feedback (loss) from the target model is used to iteratively update the weights of the surrogate models and the frequency band weights within the FPG, refining the attack's effectiveness.

Datasets

  • NeurIPS-17 dataset: A subset of 1000 images was used for evaluation. This dataset is described as containing high-quality images similar to those in ImageNet, with detailed category and target labels.
  • ImageNet: The target models (VGG-19, DenseNet-121, ResNeXt-50) were pre-trained on ImageNet. The surrogate models were also pre-trained on ImageNet.

Key Insights and Ideas

  • Frequency Domain is Key: The paper strongly suggests that manipulating specific frequency bands of an image can be a highly effective way to fool DNNs, offering a more targeted and efficient attack than manipulating the image in its entirety.
  • Low Query Count is Achievable: The most significant insight is that by combining frequency decomposition, ensemble learning with surrogate models, and adaptive weight updates, it's possible to achieve very high attack success rates with an astonishingly low number of queries (1-2 per image). This dramatically reduces the risk of detection.
  • Adaptive Weighting Matters: The adaptive adjustment of weights for both the surrogate models and the frequency bands is crucial for optimizing the attack's efficiency and success rate.
  • Trade-off Between Computation and Stealth: The method requires significant computational resources for generating perturbations on surrogate models, but this is done offline. The critical advantage is the minimal interaction with the target model, making the attack stealthy.

Key Results

  • High Attack Success Rate (ASR):
    • Targeted Attacks: Achieved over 98% ASR on VGG-19, DenseNet-121, and ResNeXt-50.
  • Non-Targeted Attacks: Achieved nearly 100% ASR on these models.
  • Minimal Average Query Count (AQC): The method consistently achieved an AQC of 1.0 for non-targeted attacks and 1.1-1.2 for targeted attacks across the tested models. This is a significant improvement over state-of-the-art methods that require tens or even hundreds of queries.
  • Robustness Against Defenses: The proposed method showed a less pronounced reduction in ASR when attacking models equipped with adversarial defense mechanisms compared to other methods.
  • Effectiveness in Frequency Domain: Comparative experiments showed that attacks in the frequency domain were more efficient, requiring fewer queries than image-space attacks.
  • Visualization: Class Activation Maps (CAMs) of adversarial examples showed a significant shift in the model's focal areas, indicating effective deception.

Limitations and Future Work

  • Computational Resources: The method requires significant computational resources due to the large number of surrogate models used. However, this computation is performed offline and does not involve the target model.
  • Time-Consuming Surrogate Model Attacks: Generating perturbations on surrogate models can be time-consuming (around 50 seconds per query), leading to longer overall attack durations.
  • Future Work:
    • Investigate the specific effects of different frequency bands on attack success.
  • Develop adaptive defenses that can counter adversarial perturbations across various frequency ranges.
  • Extend the validation to UAV-specific imagery datasets.
  • Explore applications in dynamic or resource-constrained contexts.

Summary updated

Abstract

In recent years, unmanned aerial vehicles (UAVs) vision systems based on deep neural networks (DNNs) have made remarkable advancements, demonstrating impressive performance. However, due to the inherent characteristics of DNNs, these systems have become increasingly vulnerable to adversarial attacks. Traditional black-box attack methods typically require a large number of queries to generate adversarial samples successfully. In this paper, we propose a novel adversarial attack technique designed to achieve efficient black-box attacks with a minimal number of queries. We define a perturbation generator that first decomposes the image into four frequency bands using wavelet decomposition and then searches for adversarial perturbations across these bands by minimizing a weighted loss function on a set of fixed surrogate models. For the target victim model, the perturbation images generated by the perturbation generator are used to query and update the weights in the loss function, as well as the weights for different frequency bands. Experimental results show that, compared to state-of-the-art methods on various image classifiers trained on ImageNet (such as VGG-19, DenseNet-121, and ResNext-50), our method achieves a success rate over 98% for targeted attacks and nearly a 100% success rate for non-targeted attacks with only 1–2 queries per image.
BibTeX
@inproceedings{zhang2024efficient,
  author = {Zhun Zhang and Qihe Liu and Shijie Zhou and Wenqi Deng and Zhewei Wu and Shilin Qiu},
  title = {{Efficient Ensemble Adversarial Attack for a Deep Neural Network (DNN)-Based Unmanned Aerial Vehicle (UAV) Vision System}},
  booktitle = {Drones},
  year = {2024},
  doi = {10.3390/drones8100591},
}
Figure