XFedRec

BEng (Hons) Software Engineering | Final Year Project

Keshav Ravichandran

Problem: Modern streaming platforms rely on accurate rating prediction systems to suggest media, but centralizing user interaction data poses severe privacy risks. Federated Learning (FL) mitigates this by keeping data on local edge devices, but standard FL frameworks operate under two flawed assumptions: that user preferences remain static over time, and that all participating edge devices are honest. Methodology: This project proposes a robust Federated Neural Rating Prediction framework evaluated on the MovieLens-100K dataset. The system utilizes a Multi-Layer Perceptron (MLP) architecture trained locally via Mean Squared Error (MSE). To address temporal preference shifts, a hybrid 1D Kalman Filter and ADWIN drift detector is implemented to monitor the continuous validation loss stream, triggering localized adaptation protocols when concept drift occurs. To defend against data poisoning, a Coordinate Median aggregator replaces standard FedAvg to filter out malicious Byzantine updates. Finally, local LIME integration is deployed as a proof-of-concept for edge-device Explainability. Results: Experimental results demonstrate that standard FedAvg suffers a catastrophic gradient collapse under a 20% Byzantine scaling attack. In contrast, the proposed system successfully shields the global model and autonomously recovers from synthetic concept drift, driving the Root Mean Squared Error (RMSE) down to 2.00. Subject Descriptors: • Computing methodologies → Machine learning → Machine learning approaches → Neural networks • Security and privacy → Human and societal aspects of security and privacy → Privacy-preserving protocols