AIML Research Seminar: Adversarial attacks against machine learning models for satellite imagery

Abstract: Satellites utilise a wide range of sensing modalities to collect data for applications such as earth observation, navigation and disaster management. Hyperspectral imagery is one such modality that captures a detailed spectral response of scenes across the electromagnetic spectrum. This type of imagery, however, generates vast amounts of data which requires processing to extract meaningful insights. To permit real-time decision making, edge computing – performed directly onboard a satellite – is used to process data before transmission to ensure only essential data is downlinked, ensuring fast response times. This is critical for time-sensitive applications including disaster detection and military surveillance. Deep neural networks (DNNs) are increasingly being used to automate this processing onboard satellites due to their ability to detect objects and anomalies faster and more accurately than traditional methods. However, DNNs are susceptible to adversarial attack. These attacks involve placing engineered objects into the scene of an image that cause deep neural networks to produce incorrect outputs. This research explores the vulnerability of DNNs to adversarial attack in satellite imagery and investigates methods to enhance the reliability of DNNs in the presence of such threats.