Susceptibility & defense of satellite image-trained convolutional networks to backdoor attacks

With growing opportunities and incentives to disrupt critical space infrastructure systems, cybersecurity of data pipelines associated with satellite images is becoming an increasingly important task. Combined with this threat is a growing body of literature focused on data poisoning and backdoor attacks on deep learning models tailored toward image recognition. To date, there is no work exploring backdoor attacks on deep learning models trained on satellite data. In this study, we evaluate the susceptibility of convolutional networks trained on satellite images to a common trigger-based backdoor attack. We first insert the attack into models trained on two satellite datasets and assess the susceptibility of each model to the attack. We then explore the capability of two state-of-the-art countermeasure tech-niques based on neuron activity to detect and repair the damage inflicted by the attack. Our results show significant and highly varied attack susceptibility ranging from 0 to 100% with strong dependence on the combination of classes targeted and features common in satellite images. Existing defense strategies are shown to be generally ineffective for this style of attack on satellite data. Finally, we introduce a diagnostic approach for the detection of poi-soning vulnerability in satellite image-based deep learning approaches.(c) 2022 Elsevier Inc. All rights reserved.

Automated summary

With increasing opportunities to disrupt critical space infrastructure systems, ensuring the cybersecurity of data pipelines associated with satellite images has become crucial. This study evaluates the vulnerability of convolutional networks trained on satellite images to trigger-based backdoor attacks and explores countermeasure techniques to detect and repair the damage caused by such attacks.