Defense Against False Data Injection and Fault Tolerance Optimization for Resource Scheduling in Software-Defined Satellite-Terrestrial Networks Using SFT-LoRA

Abstract

With the widespread application of Satellite-Terrestrial Integrated Networks (STINs) in key fields such as disaster monitoring, military communications, and the Internet of Things, their resource scheduling systems are facing severe security threats. This paper proposes a Software-Defined Flexible Topology SFT training mechanism that combines spatiotemporal feature analysis and a dynamic fault-tolerant scheduling optimization model. Firstly, in the detection layer, a cross-layer feature extractor based on bidirectional LSTM-CNN is designed, and SFT training mechanism is adopted to realize real-time recognition of injected data; Secondly, at the fault-tolerant layer, build an elastic resource reallocation strategy, predict task demand fluctuations through sliding windows, and dynamically adjust LoRA weight priority and node redundancy; The models were trained and validated on a satellite-terrestrial simulation dataset generated with NS-3 and OpenStack, including 100 nodes, 12 typical attack types, and over 12,000 samples. The training process integrated bidirectional LSTM layers (128 units), CNN filters, and reinforcement learning with PPO for adaptive scheduling. Finally, the experiment is based on NS-3 simulation platform and OpenStack cloud environment, and the injected attack samples cover 12 typical attack modes. Results show that the proposed framework outperforms traditional IDS and a defenseless baseline, achieving a 98.5% FDI detection rate with a false alarm rate below 1.2%. Moreover, the task completion rate after attacks increases to 96.3%, scheduling delay is reduced by 21.7%, and resource utilization is improved by 18.4%. This framework provides an effective solution for resource scheduling security in highly dynamic and weakly connected environments.