Communication-Efficient Personalized Federated Learning via Layer-Wise Multi-Threshold Random Sketching

The researchers proposed a highly efficient personalized federated learning framework based on hierarchical multi-threshold random sketches. This approach assigns independent sets of quantization thresholds to each layer of the neural network, enabling compressed representations to adapt to the specific statistical characteristics of each layer and using multiple intervals to provide fine-grained low-bit descriptions. Experimental results show that this method supports two-way communication and uses compact low-bit sketches, significantly improving the balance between communication overhead and model accuracy compared to existing single-threshold one-bit compression methods.