Study Finds Remote Sensing AI Models Vulnerable to Physical and Digital Attacks
A study by researchers from Northwestern Polytechnical University and The Hong Kong Polytechnic University found that deep learning models used in remote sensing for object detection and image classification are vulnerable to both physical and digital attacks. The vulnerabilities could affect critical applications including environmental monitoring, disaster management and autonomous transportation.
What happened
Researchers from Northwestern Polytechnical University and The Hong Kong Polytechnic University conducted a review of existing research on deep learning models used for object detection and image classification in remote sensing. The study, led by Shaohui Mei and published in the Journal of Remote Sensing on September 11, 2024, assessed the robustness of various DL model detectors including YOLO versions, RetinaNet, and FreeAnchor.
The analysis revealed vulnerabilities in deep learning algorithms for object detection that could be exploited by attackers. The researchers tested the effects of natural noise sources including salt-pepper and random noise, rain, snow and fog at different intensities, as well as adversarial attacks such as the Fast Gradient Sign Method (FGSM). The study also examined potential physical manipulations, such as attaching patches to objects to deceive AI systems.
Reported impact
The study found that AI models used in remote sensing applications are susceptible to both physical and digital attacks. These vulnerabilities could impact critical applications such as environmental monitoring, government intelligence, autonomous transportation, urban planning and disaster management.
Index analysis
The research demonstrated that training an adversarial attack is similar to training a neural network and is influenced by factors like training data and optimization strategies. Deep learning models, despite their powerful capabilities in remote sensing applications, are susceptible to different kinds of disturbances, including adversarial attacks.
Governance controls
The researchers emphasized the importance for developers and users of remote sensing technologies to be aware of these vulnerabilities and to work towards improving model robustness to ensure reliable performance in real-world conditions.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.