Hierarchical Selection-Based Local Differential Privacy Federated Learning for Distributed Robotic Vehicle Collaboration with 6G

With the advancement of 6G communication technology, the application of local differential privacy (LDP) in federated learning (FL) for the distributed collaboration of robotic vehicles has emerged as a promising solution. However, in high-dimensional models, a trade-off exists between performance and privacy protection when using LDP. Most existing research assumes that all model layers are equally important, uniformly adding random noise to the gradient of each layer. This approach leads to excessive noise injection in high-dimensional models, significantly degrading the accuracy of the global model. To address this issue, we propose SHLDP, a hierarchical differential privacy federated learning (DPFL) method. SHLDP quantifies the importance of local model layers using the Fisher information matrix and allocates a portion of the privacy budget to randomly select layers based on their importance. Layers deemed more critical are assigned a higher probability of being selected. Random noise is then added only to the selected layers. We formally prove the privacy guarantees of SHLDP and validate its effectiveness through experiments conducted on three publicly available datasets with varying data distributions. The experimental results demonstrate that SHLDP significantly enhances model accuracy while maintaining equivalent privacy constraints, e.g., in a federated learning scenario with 10 clients training on the MNIST dataset, SHLDP achieves a 1.43% accuracy improvement over FedDPA.