FedPerKD privacy aware personalized federated learning with knowledge distillation for edge oriented tomato leaf disease detection
TL;DR
FedPer-KD, a privacy-aware federated framework that combines personalization with client-side knowledge distillation, is presented, which is effective under mild client heterogeneity, and is edge-oriented for downstream deployment.
Abstract
Tomato leaf diseases threaten global food security by reducing crop yield and quality. Centralized deep-learning detectors require farms to share raw images, which raises privacy, bandwidth, and data-sovereignty concerns. We present FedPer-KD, a privacy-aware federated framework that combines personalization with client-side knowledge distillation: each client keeps its raw images and personalized classifier head locally while the shared MobileNet-V2 feature extractor is aggregated, and a local EfficientNet-B0 teacher supervises the student through softened logits. We evaluate on the PlantVillage tomato leaf dataset (18,345 images, ten classes) using simulated federated clients ( N = 2, 4, 6) trained for 10, 20, and 30 rounds with five random seeds. FedPer-KD reaches 99.34% test accuracy and 99.33% macro-F1 at six clients and 20 rounds while transmitting only the shared base (~ 8.9 MB per client per round, a 36% saving over FedAvg on the same student). Across 2–6 clients it consistently outperforms FedAvg, FedProx, Scaffold, FedPer, FedAvg + KD, and a capacity-matched FedAvg baseline; paired statistical comparisons with Holm-Bonferroni correction confirm significance ( p < 0.05) and Cohen’s d > 1.0 on every comparison. We also report ablations over the Dirichlet heterogeneity parameter (alpha = 0.5, 0.3, 0.1), the teacher fine-tuning schedule, the distillation temperature tau and weight alpha, and a battery of synthetic image corruptions. The framework is privacy-aware rather than formally privacy-preserving, is effective under mild client heterogeneity, and is edge-oriented for downstream deployment.
