Space-Borne Machine Learning
Introduction
The Sentient Satellites Lab (SSL) is helping to foster the concept of AI-enabled satellite swarms, where the long term goal is to develop a network of satellites that work cooperatively towards achieving a common mission, such as disaster prevention and mitigation, space debris monitoring, and large-scale environmental monitoring. Relative to existing space utilisation approach that employs a single/monolithic satellite for each mission, cooperative satellite swarms improve the agility and resilience of the mission.
Operationalising ML in space
AI-enabled satellite swarms require robust inter-satellite communication and a high degree of intelligence on the satellites. SSL is actively contributing to the latter research thrust, through partnerships with the space industry. Specific research activities include efficient machine learning on satellite-borne compute payloads (e.g., neural network accelerators and co-processors) and alleviating the constraints of running machine learning models long term on a space-based platform.
Relevant publications
Our group investigates fundamental and practical questions of executing machine learning models on space-borne hardware, such as the Cognisat XE-1 co-processor (pictured above) which conains an Intel Myriad 2 neural network accelerator.
Adversarial resilience of space-borne machine learning models
Deep neural networks (DNNs) have become essential in processing (both onboard and downstream) the vast quantities of data acquired by Earth-observing (EO) satellites. However, the vulnerability of DNNs towards adversarial examples, i.e., carefully crafted inputs aimed at fooling the models into making incorrect predictions, is well documented. The danger posed by adversarial attacks against machine learning algorithms for space applications is therefore of significant concern. At SSL, we are researching DNN-based machine learning algorithms that are robust against adversarial attacks. First, to illustrate the risk of adversarial attacks, we have demonstrated effective physical adversarial attacks against DNNs for aerial imagery, whereby adversarial patterns optimised by an algorithm is placed on a target object (cars) to evade detection from a DNN. Another major aspect of the research is to develop DNNs that are robust against such adversarial attacks, not only for the aerial object detection task, but also the semantic segmentation task which power valuable applications such as land-use classification, post-disaster assessment, and environmental monitoring.
Relevant publications
Adversarial pattern installed on a car to evade detection by a DNN for processing aerial imagery.
Adversarial pattern installed off and around a car to evade detection by a DNN for processing aerial imagery.