Jinhai Zhang

Privacy-preserving decentralized federated learning (DFL) has been extensively studied to protect the privacy of both data and models. Meanwhile, researchers have developed various efficient aggregation rules to defend DFL against poisoning attacks from Byzantine adversaries. However, existing techniques fail to effectively address the coexistence of model privacy and Byzantine robustness in decentralized scenarios. To bridge this gap, we propose BP-DFL, a novel DFL scheme that simultaneously detects malicious models while preserving model privacy. Our approach integrates functional secret sharing and multi-party homomorphic encryption schemes to design a secure model verification and aggregation protocol. Unlike existing solutions based on differential privacy, our scheme avoids noise injection and ensures model consistency throughout the protocol. Furthermore, we ensure that clients remain unaware of which neighbors' models are aggregated or discarded, thereby preventing clients from compromising the privacy of neighbors' models through selective aggregation. We provide formal security proofs for BP-DFL against Byzantine clients. Experimental results show that existing aggregation rules suffer from poor performance in defending against Byzantine poisoning attacks with the introduction of differential privacy noise, whereas BP-DFL achieves higher and more stable model accuracy by a substantial margin.

Graph neural networks (GNNs) are extensively employed in location-related scenarios, relying on aggregation to gather features from neighboring nodes based on edge weights. Features are closely bound to nodes’ locations and edge weights mirror distance correlations. In this sense, certain privacy concerns exist while providing location-based services if there is insufficient privacy protection. To this end, we propose a privacy-preserving and distance-aware data aggregation framework (PEDA) for GNNs. Specifically, PEDA achieves location privacy by combining circular-based positional coding with inner product functional encryption. Because of the masks in the codes, the decryption returns masked distances, preventing distance leakage. Following this, in order to protect feature privacy, we employ secret sharing. To preserve the collection strategy’s privacy, we implement an oblivious transfer for collecting the shared features. Additionally, we securely generate the adjacency matrix and aggregate features based on multi-party computation. Thorough security analysis and comprehensive evaluation demonstrate the privacy, feasibility and practicality of our approach. When compared to related works, PEDA offers four types of privacy, maintains distance awareness and feature utility, and allows for oblivious data collecting with little computational cost sacrifice.

Privacy-preserving federated learning framework has been widely applied in the Internet of Vehicles (IoV) scenario, to enhance the privacy of local models and sensitive vehicle data. However, as increasing cyberattacks are launched against both the RSU and server side, such as node compromising and data inference attacks, it is necessary to improve the robustness of uploading local updates and prevent servers from obtaining additional private information of vehicles in the two-layered IoV model, while existing schemes hardly meet all the above-mentioned security requirements. Therefore, we propose a robust and unlinkable privacy-preserving federated learning framework RUPT-FL, that tolerates dropout RSUs while disassociating local updates from the RSUs who gathered them. Based on packed secret sharing, secure multi-party computation, and homomorphic encryption techniques, we design a robust and lightweight multi-RSUs-aided local update upload protocol for vehicles, then construct an identity-oblivious update reconstruction and aggregation protocol under the two-server model. We conduct a series of experiments to demonstrate that the proposed framework not only improves the robustness and accuracy of federated learning against a semi-honest adversary, but also avoids expensive computational overhead for vehicles.