The increasing complexity and decentralization of modern blockchain networks have highlighted the limitations of traditional consensus protocols when operating under adverse or dynamic conditions. Existing approaches often fail to adapt to real-time anomalies such as Sybil attacks, network congestion, or node failures, resulting in decreased throughput, increased latency, and reduced security. Furthermore, most systems lack autonomous mechanisms to adjust operational policies based on context, especially in edge computing environments where resource constraints and topological variability demand flexible and efficient solutions. This work proposes an adaptive consensus architecture that integrates a graph-based Proximal Policy Optimization (PPO) reinforcement learning agent capable of detecting malicious behavior, optimizing validation paths, and dynamically modifying consensus logic in response to adversarial scenarios. The model is trained on a hybrid dataset composed of real traffic traces and synthetically generated adversarial behaviors, and evaluated in stress-testing environments with multiple threat vectors. Experimental results demonstrate that the proposed system maintains stable throughput (TPS) while reducing average consensus latency by 34% relative to baseline protocols under adverse high-load conditions. Regarding security, it achieves high detection in Sybil and node-collapse scenarios (DR exceeding 0.90 with FPR below 0.10), and moderate detection under congestion and erroneous transactions (DR between 0.58 and 0.70, FPR between 0.14 and 0.22). Additionally, we observe up to 16% lower average energy consumption in high-congestion settings. Energy consumption is reduced by up to 17% in crash-prone scenarios. The architecture demonstrates stable convergence over 100 operating cycles and robust adaptation to topological changes, validating its applicability in real-world deployments.
Gutierrez et al. (2025) studied this question.