Enhanced SageAttention2-Based GNNs for Reliable Block Propagation Optimization in Blockchain-Enabled Internet of Agents

Abstract

Internet of Agents (IoA) is an emerging paradigm in which autonomous Artificial Intelligence (AI) agents collaborate across organizational boundaries. Since distributed agent interactions lack a pre-established trusted central authority, blockchain, as a decentralized technology, can be integrated into IoA for decentralized and tamper-proof trust management, where service interaction records among agents are recorded into blocks through reliable consensus mechanisms. However, inefficient block propagation during the consensus process can cause redundant and outdated information. Moreover, existing block propagation schemes cannot handle dynamic block sizes caused by varying AI agent workloads or detect strategic attacks, where malicious miners strategically alternate between honest and malicious behaviors. To this end, in this paper, we propose a reliable block propagation optimization framework for public blockchain-enabled IoA. Specifically, we first design a metric called Dynamic Age of Block (DAoB) to measure block freshness under heterogeneous AI agent workloads. Then, we propose a four-element subjective logic model that incorporates behavioral volatility into ternary subjective logic models to rapidly detect changes in miner behavior. Furthermore, we develop an enhanced SageAttention2-based Graph Neural Network (GNN) model, in which quantized attention is adapted from parallel inference to sequential reinforcement learning, thereby facilitating efficient and reliable block propagation optimization. Numerical results demonstrate that the proposed model outperforms baseline methods, achieves up to 84.8% memory reduction compared with standard graph attention, and achieves an average 36% reduction in DAoB-based block propagation latency under low-bandwidth conditions compared with the gossip protocol.