Content area
Federated Learning (FL) is a distributed machine learning approach that enables collaborative model training without centralizing raw data, promising privacy but having been shown to be vulnerable to membership inference and reconstruction attacks. Homomorphic Encryption (HE) can address those privacy concerns by allowing computation on encrypted updates. However, its high computational cost and ciphertext expansion hinder deployment in resource-constrained environments. Hybrid Homomorphic Encryption (HHE) offers a more practical alternative by combining lightweight symmetric ciphers with HE, reducing client overhead while maintaining end-to-end privacy. This thesis investigates the application of HHE in FL and proposes a novel HHE-based FL framework that combines the PASTA symmetric cipher with the BFV HE scheme. Clients encrypt local model updates with PASTA and send both the lightweight ciphertext and the BFV encryption of the PASTA key to the server. The server then performs a homomorphic evaluation of the decryption circuit of PASTA and aggregates the resulting BFV ciphertexts. To mitigate confidentiality risks stemming from the use of a shared HE key, two mitigation strategies were developed: RSA wrapping, which re-encrypts the BFV-encrypted PASTA key under the server's RSA public key, and Masking, where clients encrypt a masked key that the server later unmasks during aggregation. The system was integrated into the Flower framework and evaluated under an independent and identically distributed partitioned version of the MNIST dataset with 12 clients across 10 training rounds. The results demonstrate that the HHE schemes achieved accuracy comparable to plaintext training (97.58\%-98.33\% vs. 98.93\%), while reducing the encrypted model size by a factor of 2,077$\times$, client upload traffic by up to 61.73×, and cutting client runtime by at most 30\% compared to a system based solely on the BFV HE scheme. However, server computational cost increases by roughly 6700$\times$ for each client participating in the training phase. Overall, the study shows that HHE provides an effective balance between privacy, efficiency, and accuracy in FL for resource-constrained devices. While server-side scalability remains a challenge, the proposed approach constitutes one of the first practical demonstration of HHE for FL.