SMAPPO: A security-aware multi-agent reinforcement learning framework for secure computation offloading in SAGIN

With the deployment of the 6G space-air-ground integrated network, IoT devices face challenges such as limited computational capacity, insufficient energy, and security risks. To address these issues, this paper proposes a security-aware multi-agent reinforcement learning framework for secure computation offloading ⁠—SMPPO. The approach combines Proximal Policy Optimization with an adaptive clustering algorithm and utilizes a weighted sum method for multi-objective optimization to balance security, energy consumption, and latency. By introducing a task security cost quantification model and the concept of "Security Provided," this paper quantifies the security levels under different encryption schemes. The adaptive clustering algorithm dynamically adjusts encryption schemes based on network load, ensuring data confidentiality while improving the feasibility of offloading tasks. Experimental results demonstrate that the proposed method effectively reduces task drop rates while balancing energy consumption and latency, validating its effectiveness in enhancing both security and resource efficiency.