A Noise-Aware Federated Learning Framework for Fish Spoilage Prediction in Small-Scale Fisheries: An Architecture-Controlled Performance Evaluation

Abstract

Post-harvest fish spoilage remains a major challenge for small-scale fisheries in developing economies, with losses estimated at 25-30% attributed largely to inadequate cold chain monitoring at informal landing sites and markets. This paper presents a noise-aware Federated Learning (FL) framework for predicting fish spoilage risk, evaluated through an architecture-controlled comparison against a centralised baseline. Because a large, labelled dataset from operational cold chain deployments was unavailable, a synthetic dataset of 10,000 records was generated using environmental ranges reported in the literature and calibrated to reflect DHT22 sensor tolerances, explicitly separating true environmental conditions from sensor-reported values to model realistic calibration drift and measurement noise. An identical Multi-Layer Perceptron (MLP) architecture was used for both the centralised baseline and the federated system, implemented using the Flower framework across three simulated nodes representing distinct fisheries sites, ensuring that any observed performance difference reflects the federated training process itself rather than a difference in model capacity. The centralised model achieved 94.75% accuracy, within half a percentage point of a calculated 94.73% theoretical noise ceiling derived from the dataset’s own noise model, while the federated model achieved a comparable 94.8-95.0% accuracy across five training rounds. A STRIDE-based threat analysis identified nine cybersecurity threats affecting the system, of which the highest-severity communication threat was mitigated using Transport Layer Security (TLS), implemented and empirically measured to add negligible computational overhead. Permutation-based feature importance analysis further identified exposure duration, rather than temperature, as the dominant predictor of spoilage risk (74.48% relative contribution). These results demonstrate that federated learning’s privacy-preservation and connectivity-resilience benefits can be obtained at negligible measurable cost to predictive accuracy, supporting federated architectures as a practically competitive, rather than purely privacy-motivated, choice for connectivity-constrained fisheries contexts in developing economies.