Enhancing Edge Multipath Data Security Offloading Efficiency via Sequential Reinforcement Learning

The multipath transmission structure decouples network services from a single transmission carrier, which has great potential for shaping a more secure and efficient 6G network. Existing multipath transmission schemes face challenges such as network heterogeneity, perception lag, and additional scheduling delay, which limits their ability to improve bandwidth aggregation capacity and information security. To address these issues, we propose the Sequential Reinforcement Evolution (SRE) scheme, which utilizes deep reinforcement learning to predict the value of future scheduling actions based on past network states. The SRE scheme regards improving bandwidth aggregation capacity and anti-eavesdropping ability as optimization goals, and designs a semi-symmetric attention recurrent neural network (SARNN) to better mine the sequential nature of the scheduling process. The SRE scheme utilizes approximately 500 million real network data points to pre-train the SARNN model, and performs cycle optimization during the actual deployment process. Experimental results show that SRE significantly outperforms state-of-the-art scheduling schemes with a 32% increase in bandwidth aggregation and a 117% increase in traffic security dispersion with minimal impact on latency.