Optimized Consensus with DAGWise: A GNN-Enhanced Approach for Scalable and Fault-Tolerant DAG-Based BFT

DAG-based Byzantine Fault Tolerant (BFT) protocols improve scalability by enabling parallel block proposals and decoupling data dissemination from consensus. However, they still rely on static leader selection rules and incur high communication overhead, particularly under adverse network conditions. We propose DAGWise, a learning-augmented consensus framework that integrates Graph Neural Networks (GNNs) and Structural Equation Models (SEM) to infer validator dependencies and prioritize commit decisions. DAGWise computes structure-aware leader scores from local DAG views, enabling selective quorum evaluation and adaptive message pipelining. This enhances latency, throughput, and robustness without modifying underlying quorum logic. Evaluations on geo-distributed deployments show that DAGWise achieves up to 2× lower commit latency and 1.5× higher throughput compared to baseline DAG-BFT protocols under both normal and adversarial settings.