Byzantine-resilient federated learning with dynamic scoring matrix and variant PBFT consensus under differential privacy

The increasing concerns regarding data privacy exacerbate the challenges associated with “data silos”. Federated learning (FL) effectively addresses these issues by facilitating distributed machine learning without necessitating direct data exchange. However, the dependence on a central server in conventional FL architectures exacerbates privacy risks and limits cross-domain data sharing. Existing blockchain-based FL frameworks often employ static consensus protocols, such as classical Practical Byzantine Fault Tolerance (PBFT), which typically rely on fixed weight aggregation strategies. While these methods simplify implementation, they fail to adaptively adjust aggregation weights according to heterogeneous privacy budgets. Attempts to implement adaptive weight aggregation often require achieving consensus for each individual weight, significantly reducing efficiency and creating scalability challenges in large-scale networks. To address these gaps, we propose DSM-PBFT, a variant PBFT consensus enhanced with dynamic scoring matrices (DSM), which enables parallelized validation of multiple models while adaptively adjusting aggregation weights based on differential privacy budgets. Our noise-aware aggregation mechanism dynamically reweights models through cross-validation of accuracy, F1 score, and loss-transformed metrics, effectively decoupling privacy guarantees from model utility degradation. Security analyses affirm the robustness of this framework against Byzantine attacks, with experimental results on MNIST, FashionMNIST and CIFAR-10 demonstrating superior model accuracy across diverse privacy budgets while effectively curbing accuracy degradation under attack scenarios.