EHFL: Efficient Horizontal Federated Learning With Privacy Protection and Verifiable Aggregation

As a recently distributed machine learning framework, federated learning (FL) has garnered attention for its privacy protection. However, recent researches in recent years have shown malicious entities may still acquire the clients’ privacy in FL. Moreover, factors, such as the verifiability of the aggregation model and the huge computational and communication overheads, also make existing solutions less practical. To this end, we design a novel FL architecture named EHFL. First of all, EHFL protects data privacy by concealing the clients’ data using a single mask and group key encryption. Second, combined with symmetric balanced incomplete block design (SBIBD), our EHFL dramatically reduces client computational and communication overhead to approximately $(k+1)/(k^{2}+k+1)$ compared to traditional FL (e.g., FedAvg). Third, EHFL designs an ingenious verification mechanism to ensure the correctness of the aggregation server’s results via the Hamiltonian graph formed by the SBIBD. Finally, sufficient theoretical analyses prove the reliability of EHFL and lots of experiments demonstrate the effectiveness of EHFL.