ALPPA: An accuracy-lossless and privacy-preserving aggregation strategy for federated knowledge graph completion model

Federated knowledge graph completion facilitates data sharing and collaborative modeling among multiple clients, providing an effective solution to complete missing knowledge. To achieve privacy-preserving while maintaining model accuracy, researchers have proposed some aggregation strategies. However, these strategies typically focus solely on protecting the privacy of local model parameters uploaded by clients to the server, while neglecting the privacy of the aggregated global model parameters on the server. This limitation may allow the server to steal the global model parameters, posing risks of privacy leakages and potential economic losses for participating clients. To address this issue, we propose an accuracy-lossless and privacy-preserving aggregation (ALPPA) strategy. In our approach, clients inject noise into their local model parameters before uploading them to the server and use a shift matrix to disrupt the noise sequence. This ensures that the server cannot recover the original noise sequence, thereby preventing the global model parameters from being exposed to the server and achieving dual privacy-preserving for both local and global model parameters. After receiving the aggregated global model parameters, clients perform a noise removal operation to restore accurate and noise-free global model parameters. Since the noise is completely removed before the local training phase, the final federated model achieves the same accuracy as a model trained without noise, thus ensuring lossless accuracy. Building on ALPPA, we further propose an optimized strategy (ALPPA+) that significantly improves communication efficiency. Optimizing the noise injection module by compressing the noise matrix into a vector, ALPPA+ reduces the uplink communication cost by 50% compared to ALPPA. This optimization maintains privacy-preserving and model accuracy while significantly improving communication efficiency, offering a more efficient and secure solution for practical federated knowledge graph completion applications.