Robust Trust-Aware Federated Learning for Privacy-Preserving Distributed Intelligence in Resource-Constrained IoT Systems: A Blockchain-Assisted Architecture

Abstract

The rapid expansion of Internet of Things (IoT) ecosystems has intensified the need for distributed intelligence mechanisms that reduce direct raw-data exposure while remaining resilient to adversarial manipulation. Federated learning (FL) addresses part of this challenge by enabling collaborative model training without centralizing raw data, but it remains vulnerable to malicious client behavior, particularly model poisoning attacks that can substantially degrade global model quality. This paper investigates a blockchain-assisted, trust-aware FL framework for privacy-aware distributed intelligence in resource-constrained IoT systems, where the blockchain layer is used to support trust coordination, auditability, traceability, and tamper-resistant metadata recording rather than to directly improve predictive performance. The empirical study compares six aggregation strategies: FedAvg, coordinate-wise Median, Trimmed Mean, FLTrust, trust-aware weighted aggregation, and a hybrid trust-trimmed mean method. The primary evaluation is conducted on the UCI Human Activity Recognition (UCI HAR) dataset under Dirichlet-based non-IID client partitioning α = 0.1 and α = 1.0, partial client participation, and sign-flip model poisoning. Each configuration is evaluated over five independent runs. Under the severe 40% malicious-client stress test, FLTrust achieves the strongest mean robustness among the evaluated methods, reaching 0.4195 ± 0.1295 accuracy and 0.3108 ± 0.1294 macro-F1 for α = 0.1, and 0.6472 ± 0.0725 accuracy and 0.6014 ± 0.0949 macro-F1 for α = 1.0. In lower-intensity attack controls with 10% and 20% malicious clients, the trust-aware and hybrid trust-trimmed strategies are the most competitive, achieving the highest or near-highest mean performance without requiring a clean server-side reference set. Mechanism-level analysis shows that FLTrust is particularly effective in the severe setting because its reference-based scoring assigns near-zero weights to malicious clients, whereas trust-aware and hybrid methods rely on relative update consistency and are more affected by the interaction between poisoning and statistical heterogeneity. Globally, the results indicate that no single aggregation rule dominates all adversarial regimes. Instead, reference-based trust provides strong protection under severe poisoning, while trust-aware and hybrid aggregation offer competitive root-free alternatives under lower attack intensities. The findings also clarify that the privacy-preserving scope of the framework derives from FL-based data locality and does not constitute a formal cryptographic privacy guarantee.