Workshop: 13th Annual International Workshop on Innovating the Network for Data Intensive Science — INDIS
Authors: Mohammad Arafath Uddin Shariff and Byrav Ramamurthy (University of Nebraska-Lincoln)
Abstract: Research and education networks, regional ISPs, and campus networks hold telemetry worth modeling jointly, but non-disclosure agreements and privacy policy prevent pooling it. We evaluate federated representation learning on real telemetry from ten Internet2 backbone routers and ten CESNET3 academic institutions, with a commercial ISP held out as an unseen-network probe, comparing centralized, federated, and local-only training at equal budget across five seeds. Federation is statistically equivalent to centralized pooling (5% equivalence bound, p = 0.020), so the data-custody constraint costs nothing in representation quality. Its benefit is not uniform: backbone clients gain a median 9.2% reduction in reconstruction error, campus clients 1.9% (p = 0.0003), an effect robust to outlier removal and training-set size that tracks a difference in traffic regularity we characterize but do not isolate. Two manipulations of local optimization selectively suppress the backbone benefit, identifying local adaptation as the mechanism.
Back to 13th Annual International Workshop on Innovating the Network for Data Intensive Science — INDIS Archive Listing
Back to Full Workshop Archive Listing