Joint Optimization of Utility and Privacy for Space-Air-Ground Integrated Network Task Offloading: a Co-D3QN Approach

The Space-Air-Ground integrated network (SAGIN) offers flexible and complex communication support for regions lacking infrastructure. In SAGIN, ground users (GUs) can offload computational tasks to base stations (BSs) or Low Earth Orbit satellites (LEOSs) with the assistance of uncrewed aerial vehicles (UAVs), thereby enhancing their network service experience. During the offloading process, it is essential to optimize both the latency and energy consumption of devices, as well as to preserve privacy. Considering the utility and privacy of task offloading in SAGIN, we propose cooperative dueling double deep Q-Network (Co-D3QN), an intelligent offloading algorithm based on deep reinforcement learning (DRL). Specifically, Co-D3QN first employs the Laplace mechanism (LM) to provide a strict $\epsilon$-differential privacy ($\epsilon$-DP) guarantee for user location in SAGIN, obfuscating user location and defining privacy cost. We then formalize the joint decision-making process as a Markov decision process (MDP) and design a D3QN to learn optimal actions. The proposed algorithm adopts the centralized training and decentralized execution (CTDE) architecture, utilizing long short-term memory (LSTM) during execution to predict current decisions, thereby enhancing network performance. Simulation results indicate that Co-D3QN significantly improves the system utility and the amount of privacy contained, providing an efficient and accurate offloading strategy for privacy-sensitive applications in SAGIN.