MUFFLe Boosts Federated Learning Efficiency with Deduplication

This paper introduces MUFFLe, a novel communication-efficient compression scheme designed for federated learning environments, particularly beneficial for edge devices. Federated learning often faces significant limitations due to the high uplink communication costs associated with transmitting model updates from numerous clients to a central server. MUFFLe addresses this by integrating generalized deduplication (GD) into the standard FedAvg pipeline. The core idea behind MUFFLe is to identify and deduplicate repeated patterns within the model update vectors. This approach results in a fixed-rate, variable-count compression scheme, meaning it can adaptively compress updates based on their inherent redundancy. By eliminating redundant information, the amount of data that needs to be transmitted is substantially reduced. Preliminary experimental results, conducted on an IID MNIST dataset with 20 clients, demonstrate MUFFLe's effectiveness. It achieved a target accuracy of 92.93% with a cumulative uplink communication of only 38 MB. This performance significantly outperforms other common compression techniques, such as 8-bit quantization (75 MB) and Top-k sparsification (86 MB), and drastically reduces communication compared to uncompressed FedAvg (310 MB). These findings highlight the viability of applying generalized deduplication to enhance communication efficiency in federated learning.