Avaliação de privacidade e eficiência em sistemas de recomendação federados com criptografia homomórfica

Abstract

Recommender systems traditionally rely on the centralization of sensitive data, generating conflicts with modern data protection regulations, such as the LGPD. Federated Learning (FL) mitigates this risk by decentralizing the training process, but shared gradients remain vulnerable to inference attacks. This work investigates the integration of Homomorphic Encryption (CKKS scheme) into a Federated Matrix Factorization model to ensure strict privacy. Using the Flower framework and the TenSEAL library, we evaluated the FedProx algorithm to handle the data heterogeneity (Non-IID) inherent to recommender

systems. Experimental results demonstrate that predictive accuracy is fully preserved under encryption, with the homomorphic approximation noise acting as an implicit regularizer that can stabilize convergence. However, the cryptographic layer imposes a high computational cost and data expansion, exposing critical trade-offs for real-time scalable deployments.