Comprehensive Review of Data Poisoning Attacks in Green Distributed Systems

Abstract

Green distributed systems (GDS) represent a critical paradigm shift in computing, aiming to minimize energy consumption and carbon footprint through techniques like energy-aware scheduling, distributed learning, and edge-cloud offloading. However, the inherent trade-off between energy efficiency and system redundancy creates a novel and enlarged attack surface for data poisoning. This paper provides a comprehensive review of data poisoning attacks specifically targeting the sustainability objectives of GDS. The study defines data poisoning in this context not merely as an accuracy degradation threat, but as an energy integrity attack designed to induce energy-inefficient behavior, such as unnecessary recomputation or suboptimal carbon-aware scheduling. The unique attack vectors, including sponge poisoning that amplifies energy consumption in neural networks, and poisoning of energy-aware federated learning models are categorized. Furthermore, the critical defense challenges, where traditional security measures often incur prohibitive energy overheads are analyzed, negating the “green” objective. Finally, key future research directions, emphasizing the need for co-optimized robustness and energy-efficiency in GDS security policies are analyzed.