Unsupervised Wavelet Scattering Autoencoders for Efficient Malicious Client Detection in Federated Coffee Leaf Disease Classification: A Kenyan Arabica Case Study.

Federated Learning (FL) enables geographically dispersed edge devices to jointly train a shared model without transmitting raw data, which is essential in precision agriculture, where bandwidth constraints, sensor heterogeneity, and data sovereignty preclude centralized processing. Yet FL remains vulnerable to malicious or faulty clients injecting corrupted data, a threat amplified in open-field deployments where nodes can be physically compromised or adversarially controlled. We address this problem in the context of Kenyan Arabica coffee leaf disease detection, using the JMuBEN 2 dataset. We propose WaSA (Wavelet Scattering Autoencoder), a fully unsupervised, proactive framework that screens clients before training begins. Each client transmits a single spectral embedding, constructed via truncated Singular Value Decomposition and a first-order Wavelet Scattering Transform, to the server once at initialization; a denoising autoencoder trained on clean auxiliary data then scores each client by reconstruction error. WaSA requires no attack labels, generalizes to unknown threat types, and adds negligible communication overhead per client. Under 90% client corruption, WaSA recovers Oracle-level accuracy (up to 100.00%), outperforming all baselines by over 17 percentage points in a regime where every existing defense collapses.