mhnarfth/who-benefits-federation: INDIS 2026 camera-ready

Abstract

Software archive accompanying the paper "Who Benefits from Federation? Heterogeneous Gains in Cross-Operator Network Traffic Representation Learning" by Mohammad Arafath Uddin Shariff and Byrav Ramamurthy (School of Computing, University of Nebraska-Lincoln), presented at the 13th International Workshop on Innovating the Network for Data-Intensive Science (INDIS 2026), held in conjunction with SC26, Chicago, 15 November 2026. The study evaluates federated representation learning across 20 heterogeneous network clients drawn from a national research backbone (Internet2, 10 routers) and a European NREN (CESNET3, 10 academic institutions), with a commercial regional ISP held out as an unseen-network probe. Each client trains an autoencoder over 24-hour traffic windows, and centralized, federated, and local-only training are compared at equal optimization budget across five seeds. Two findings. First, federated training is statistically equivalent to centralized pooling under two one-sided tests at a ±5% bound (p = 0.020), so the data-custody constraint costs nothing in representation quality. Second, the benefit of federation is not uniform: backbone clients gain a median 9.2% reduction in reconstruction error while campus clients gain 1.9% (Mann-Whitney p = 0.0003). The effect survives outlier removal and is not explained by training-set size. Two independent manipulations, proximal regularization strength and local-epoch budget, each selectively suppress the benefit for irregular-traffic clients, identifying unconstrained local adaptation as the mechanism. This archive contains the complete eight-stage pipeline, standalone diagnostic tools, every metric table and statistic underlying each figure and table in the paper, the six generated figures, and the LaTeX source of the camera-ready. Two of the three data sources are access-restricted (Internet2 NetFlow under non-disclosure agreement, and the regional ISP snapshot), so client tensors, trained checkpoints, and cached embeddings are deliberately excluded. The published metric tables allow every reported number to be verified without access to the raw data, and the pipeline accepts any equivalently formatted hourly series in their place. The public CESNET-TimeSeries24 corpus is available at doi:10.5281/zenodo.13382427.