Apollo: Permissioned Blockchain Network Auto-Tuning Based on DRL Against Eclipse Attack

Eclipse attack, one of the most common attacks in blockchain networks, not only affects the overall throughput and latency of the network, but also leads to a large number of redundant synchronization messages, resulting in a serious waste of resources. Existing studies propose auto-tuning to improve the performance of blockchains by fine-tuning the system parameters, which can also reduce the loss from eclipse attacks. However, these techniques only consider throughput and latency in the parameter search and ignore the wasted resources due to eclipse attacks, resulting in their inability to retrieve the optimal parameter configuration under eclipse attack. To address this issue, we propose Apollo, an end-to-end permissioned blockchain network auto-tuning system based on deep reinforcement learning (DRL) against eclipse attacks, and take Hyperledger Fabric as an example to verify the effectiveness of our system. The key component of Apollo is the enhanced Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm: DPB-MADDPG, where the number of agents is dynamically determined according to the number of node organizations and the type of nodes. In addition, we select two metrics to measure the resource waste of an eclipse attack based on preliminary experiments, including cluster common message sending time (CCMST) and gossip state commit duration (GSCD). In a Fabric system with 4 peers and 3 orderers, our Apollo can increase system throughput by 125.55% and reduce latency, CCMST, and GSCD by 67.24%, 67.82%, and 32.63%, respectively, compared to traditional auto-tuning.