Aviel Ben Siman Tov (Ben Gurion University of the Negev), Edita Grolman (Ben Gurion University of the Negev), Yuval Elovici (Ben Gurion University of the Negev), Asaf Shabtai (Ben Gurion University of the Negev)
Satellites’ stable operation relies on anomaly detection (AD), which is used to identify abnormal behavior in onboard systems. However, traditional AD methods struggle to function effectively in the resource-constrained environment of satellites, where energy, memory, and computation are severely limited. This challenge is especially evident in CubeSats, the most widely deployed class of small satellites, where such constraints limit the applicability of conventional AD methods and lead to a degradation in overall performance. We introduce LighTellite, a reinforcement learning-based dual-agent framework that aims to balance AD performance and energy efficiency, in which one agent determines energy budgets, and the other dynamically selects the optimal model among a pretrained pool of AD models (each with different performance and energy characteristics). LighTellite’s dynamic AD model selection enables context-aware adaptation in response to both onboard satellite data and available resources, resulting in an improvement in AD performance while maintaining low energy consumption. Experiments conducted on AegisSat, a state-of-the-art CubeSat testbed, show that our proposed framework improved attack detection rate by 10% while reducing inference energy consumption by 21.8% compared to the best static AD models (in which the same model is used throughout the entire orbit). The code and additional materials are available in the GitHub repository.
