Universal adversarial perturbation for remote sensing images

Qingyu Wang, Guorui Feng, Zhaoxia Yin, Bin Luo

Introduction

Deep learning models have become integral to Remote Sensing Image (RSI) classification, supporting critical applications in urban planning, environmental monitoring, and geographic intelligence. However, these models are susceptible to adversarial attacks—subtle, intentional perturbations that induce misclassification. While most research focuses on image-specific attacks, this study investigates Universal Adversarial Perturbations (UAPs). A UAP is a single noise pattern that, when applied to any image within a dataset, causes a target model to fail. Existing UAP methods often produce visible noise or fail to account for the unique characteristics of RSIs, such as varying resolutions and spectral data. This paper proposes a novel framework to generate highly effective and visually stealthy UAPs specifically for the remote sensing domain.

Methodology

The proposed framework utilizes a two-stage process to generate UAPs that are both powerful and difficult for human observers to detect.

The proposed two-stage framework for generating universal adversarial perturbations.

Stage 1: Training Stage

In the first stage, an encoder-decoder network is employed to learn the underlying distribution of effective perturbations. Unlike traditional methods that compute gradients for individual images iteratively, this network takes random noise as input and generates a "mid-level" perturbation. This perturbation is optimized through backpropagation to maximize the classification error of target models, such as VGG or ResNet architectures, across the training set.

Stage 2: Modifying Stage

The second stage introduces a saliency-based fine-tuning mechanism, which is the core innovation of this research. By generating saliency maps, the system identifies the specific regions of an image that the classification model prioritizes. An adjustment module then amplifies the perturbation in these sensitive areas while dampening it in less important regions. This "surgical" approach ensures high attack success rates while minimizing the overall perturbation magnitude, keeping the noise significantly less noticeable than previous methods.

Experimental Setup

The framework was evaluated using the PatternNet dataset, a large-scale RSI collection comprising 38 classes (e.g., airplane, baseball court, bridge) with 800 images per class at 256×256 resolution. The researchers tested the attack against several widely used deep learning architectures, including VGG16, VGG19, ResNet34, and ResNet101, to assess both success rates and cross-model transferability.

Results and Discussion

The proposed method demonstrated significant efficacy, achieving a peak Attack Success Rate (ASR) of 97.09% on ResNet101. Notably, the approach maintained a lower Perturbation Magnitude (PM) compared to existing state-of-the-art methods. For instance, on VGG16, the PM was reduced to 3375.28 (l2 norm), a substantial improvement over the 4126.38 reported in previous UAP studies.

Example of a misclassified remote sensing image where a basketball court is identified as an overpass.

Beyond raw success rates, the study revealed "attack selectivity," where misclassified images were frequently funneled into specific "sink" classes. For example, in VGG19, over 88% of errors were concentrated in just three classes. Furthermore, UAPs generated for one model architecture showed high transferability to others, indicating shared vulnerabilities in how different networks perceive remote sensing features. Qualitatively, the resulting adversarial images remain virtually indistinguishable from the originals to the human eye, as shown in the misclassification examples.

Conclusion

This research highlights a critical vulnerability in remote sensing classifiers. By leveraging encoder-decoder networks and saliency-based attention, the authors created UAPs that are both highly effective and visually stealthy. While the stability of saliency maps remains a challenge for future refinement, this work serves as a vital wake-up call for the remote sensing community. As these models are increasingly deployed in security-sensitive applications, addressing these universal vulnerabilities is paramount.

Summary updated

BibTeX
@inproceedings{wang2022universal,
  author = {Qingyu Wang and Guorui Feng and Zhaoxia Yin and Bin Luo},
  title = {{Universal adversarial perturbation for remote sensing images}},
  booktitle = {MMSP},
  year = {2022},
  doi = {10.1109/mmsp55362.2022.9948869},
}
Figure