An Efficient Sharing-Encryption-Based Secure Aggregation Protocol for Federated Learning in Multimedia Applications

In multimedia applications, Federated Learning (FL) has emerged as an effective training paradigm, enabling distributed clients to collaboratively train a shared model without transmitting raw data. However, FL remains vulnerable to privacy threats such as data reconstruction and membership inference attacks, which has motivated the adoption of secure aggregation protocols to protect client model updates. Existing secure aggregation schemes that combine homomorphic encryption with secret sharing predominantly follow a sharing-decryption paradigm, requiring additional client interaction during decryption and thereby incurring substantial computational, synchronization, and communication overhead. In this paper, we propose Threshold Vector Aggregation (TVA), a novel secure aggregation protocol that adopts a sharing-encryption and aggregation paradigm. Under TVA, each client encrypts its local update only once using a unique private key and uploads a single ciphertext to the server, without any further interaction. The cipher-texts are directly aggregatable yet individually undecryptable, and only the final aggregated result can be correctly recovered by the server. Extensive experiments demonstrate that TVA significantly outperforms state-of-the-art secure aggregation schemes, achieving up to 33× lower server-side aggregation time, 132× faster client-side encryption, and 170× reduction in communication overhead, while preserving model accuracy comparable to plaintext federated learning.