MUFFLe: Efficient Model Update Compression via Generalized Deduplication for Federated Learning

A new compression scheme called MUFFLe aims to cut the uplink communication cost of federated learning by applying generalized deduplication to model updates, according to a work-in-progress paper posted to arXiv on 12 June 2026 [1]. The paper, hosted on the open-access preprint repository arXiv, introduces MUFFLe as a method that integrates generalized deduplication into the standard FedAvg pipeline [1][2]. The technique identifies and removes repeated patterns across the update vector, producing what the authors describe as a fixed-rate, variable-count compression scheme [2]. Preliminary experiments used the IID MNIST dataset with 20 clients. The authors report that MUFFLe reached a target accuracy of 92.93% while consuming 38 MB of cumulative uplink communication [1][2]. By comparison, 8-bit quantization required 75 MB, Top-k sparsification used 86 MB, and uncompressed FedAvg transmitted 310 MB to achieve the same accuracy level [2]. Federated learning is often deployed in edge environments where uplink bandwidth is constrained, making the transmission of model updates a bottleneck [2]. The MUFFLe results, though described as preliminary, suggest that generalized deduplication can be applied to this problem [1]. The paper appears on arXiv as a preprint, meaning it has been moderated but not peer-reviewed [6]. arXiv, founded in 1991, now hosts more than two million articles and receives roughly 24,000 submissions per month as of late 2024 [6]. The repository’s Labs framework, which appears at the bottom of the MUFFLe abstract page, allows community collaborators to build tools such as Bibliographic Explorer and CORE Recommender that sit alongside papers [4][5]. Those tools are governed by guidelines that require partners to uphold openness, community, excellence, and user data privacy [4]. The MUFFLe paper does not include deployment benchmarks beyond the MNIST trial, and the authors frame the work as an initial feasibility demonstration [1][2]. No external validation or independent replication of the results has been cited in the preprint.