Performance Analysis of a Decentralized Federated Learning System for Spoken-Command Recognition: Resilience and Security Considerations

Abstract

In the industrial edge-to-cloud continuum, data is often privacy-sensitive and spans multiple organizations that do not fully trust one another, making central aggregation of raw data undesirable and often non-compliant with regulations such as the General Data Protection Regulation (GDPR). Federated Learning (FL) addresses this by sharing model updates rather than raw data, but conventional FL assumes a central coordinator, leaving it exposed to poisoning and inference attacks and to a single point of trust and failure. Decentralized Federated Learning (DFL) couples FL with Distributed Ledger Technologies (DLTs), removing the coordinator and enabling verifiable aggregation in trustless, cross-organizational environments. In this work, we assess the applicability of DFL to on-device spoken-command recognition—a representative edge audio task underpinning voice-driven industrial interfaces—by comparing decentralized and centralized training under idealized and adversarial conditions. Using a Convolutional Neural Network (CNN) replicated across edge nodes, we evaluate resilience to inter-node data imbalance, to poisoning attacks, and to a privacy-preserving noise-injection defense against inference attacks, together with model compression for resource-constrained edge devices. The system pairs this comparison with a validation-based poisoning defense in which each node scores its peers’ updates on its own held-out data, and an update is aggregated only if a majority of nodes report a weighted F1-score above a threshold—requiring neither a shared validation set nor a trusted validator. Our results indicate that the DFL system achieves accuracy comparable to centralized baselines in most scenarios (weighted F1-score 0.762 across nine nodes, against 0.896 centralized), and that a cross-node validation mechanism reliably excludes poisoned updates as long as fewer than half of the nodes are compromised (within 3.54% of the unpoisoned model). Noise-based inference defenses reduce accuracy substantially (44.7% on average at a noise standard deviation of 1.0), exposing a sharp privacy–utility trade-off, whereas model compression preserves performance (0.765 against 0.762 for pruning and format conversion, with 8-bit quantization costing up to a further 13.3%). These findings clarify both the promise and the current limitations of decentralized, privacy-preserving learning for the industrial edge-to-cloud continuum.