Although federated learning is gaining prevalence in smart sensor networks, substantial risks to data privacy and security persist. An improper application of federated learning techniques can lead to critical privacy breaches. Practical and effective privacy-enhanced federated learning (PEPFL) is a widely used federated learning framework characterized by low communication overhead and efficient encryption and decryption processes. Initially, our analysis scrutinized security vulnerabilities within the PEPFL framework and identified an effective attack strategy. This strategy enables the server to derive private keys from content uploaded by participants, achieving a 100% success rate in extracting participants’ private information. Moreover, when the number of participants does not exceed 300, the attack time does not surpass 3.72 s. Secondly, this paper proposes a federated learning model that integrates homomorphic encryption and secret sharing. By using secret sharing among participants instead of secure multi-party computation, the amount of effective information available to servers is reduced, thereby effectively preventing servers from inferring participants’ private gradients. Finally, the scheme was validated through experiments, and it was found to significantly reduce the inherent collusion risks unique to the federated learning scenario. Moreover, even if some participants are unavailable, the reconstructable nature of secret sharing ensures that the decryption process can continue uninterrupted, allowing the remaining users to proceed with further training. Importantly, our proposed scheme exerts a negligible impact on the accuracy of model training.

With the increasing awareness of privacy protection and the improvement of relevant laws, federal learning has gradually become a new choice for cross-agency and cross-device machine learning. In order to solve the problems of privacy leakage, high computational overhead and high traffic in some federated learning schemes, this paper proposes a multiplicative double privacy mask algorithm which is convenient for homomorphic addition aggregation. The combination of homomorphic encryption and secret sharing ensures that the server cannot compromise user privacy from the private gradient uploaded by the participants. At the same time, the proposed TQRR (Top-Q-Random-R) gradient selection algorithm is used to filter the gradient of encryption and upload efficiently, which reduces the computing overhead of 51.78% and the traffic of 64.87% on the premise of ensuring the accuracy of the model, which makes the framework of privacy protection federated learning lighter to adapt to more miniaturized federated learning terminals.