Peer-to-peer privacy-preserving vertical federated learning without trusted third-party coordinator

Federated learning is an emerging paradigm of distributed collaborative machine learning , which allows different data nodes to train together for building a better model. Recently, privacy-preserving federated logistic regression has been widely concerned by academia and industry, which exploits techniques such as secure multi-party computation (MPC), homomorphic encryption (HE), and differential privacy (DP) to train shared model among different data nodes. However, MPC-based schemes could not efficiently handle high-dimensional sparse data and have high communication burden; HE-based schemes have potential information leakage risks and high computational burden; DP-based schemes affect the speed of model convergence and cause a loss of model accuracy. Besides, some of the existing schemes require a trusted third-party (TTP) coordinator, which greatly increases the complexity of training. To address these problems, in this paper, combining the HE and secret sharing (SS), we present a peer-to-peer privacy-preserving vertical federated logistic regression (VFLR) without TTP, which can securely train a shared model over vertically partitioned large-scale data from two data nodes. Moreover, to ensure the security of the training process, the proposed scheme secretly shares model weights between two data nodes, rather than reveals them to both parties during model training. Furthermore, the proposed scheme utilizes the single instruction multiple data (SIMD) and batching to achieve parallel processing and time amortization. Finally, the experimental evaluations demonstrate that the proposed scheme has less training time and better model performance than existing schemes.