FedDSC: A Federated Learning Framework for Driving Style Profiling

Driving style recognition is essential to ensure the safety and efficiency of autonomous driving.
Current approaches to driving style analysis mainly rely on clustering to group similar driving behaviours.
However, centralised clustering methods require aggregating raw driving data from multiple sources, which compromises privacy.
They also struggle to handle the non‑IID (not independent and identically distributed) nature of real-life driving data, limiting their ability to deliver effective and robust results in real‑world scenarios.
To address these issues, we propose FedDSC (Federated Driving Style Clustering) for driving style clustering and recognition.
The FedDSC framework uses a modified autoencoder that transforms high-dimensional driving data into compact latent representations that capture key features for clustering.
Local clients independently train autoencoders on their data and share only model parameters with a central server.
The server then aggregates and updates the global model.
After federated training, the updated model is sent back to local clients for further data reduction and clustering, ensuring data privacy and effectively handling non-IID data.
Experimental results demonstrate that FedDSC not only preserves data privacy but also outperforms both local and centralised benchmark models, as well as other federated clustering methods.