Enhancing Federated Learning Security Using Homomorphic Encryption and Zero-Knowledge Proofs

Abstract

Federated learning enables multiple clients to collaboratively train a shared model without exchanging raw data, but it raises privacy and integrity concerns when model updates traverse untrusted channels. In this project, we develop a secure federated learning pipeline that combines the CKKS homomorphic-encryption scheme with Groth-16 zero-knowledge proofs to protect client updates during transmission and to verify that each update stays within an agreed-upon norm bound. We benchmark CKKS parameters (poly_modulus degree, coefficient_moduli, and scale) on real-world model vectors to identify an optimal setting—8192-degree with two primes (60-bit and 40-bit) at a 232scale—that offers sub-100 ms encryption, minimal error, and moderate ciphertext sizes ( 3.3 MB). Clients train a small convolutional network on disjoint partitions of the MSTAR SAR dataset, generate succinct ZK proofs for each 128-element weight chunk, encrypt those chunks under the selected CKKS context, and submit both ciphertexts and proofs to dedicated servers. The homomorphic-aggregation server sums encrypted updates, while the ZKP server enforces correctness by rejecting any proof that violates the norm constraint—demonstrated by catching an intentionally malicious client. End-to-end testing confirms that the combined CKKS+ZKP pipeline preserves model accuracy and ensures both confidentiality and integrity of federated updates.