Dataset and Experimental Results: Robust Aggregation Ranks the Free-Rider First

Abstract

Prototype-based federated learning exchanges class-embedding matrices without exposing the computation that produced them. We evaluate three verification mechanisms against free-riding at this interface: robust aggregation used for contribution ranking, commitments to model behavior, and passport-derived identities. On CIFAR-100, Krum assigns a client that republishes the aggregate a detection AUC of 0.011, and distances from coordinate-wise median, trimmed-mean and Bulyan outputs yield 0.000. Exact score comparisons explain this inversion, and a matched-moment counterexample shows that no criterion based on the cohort mean and scalar dispersion can predict it. A prover with no model or data passes all 40 rounds of a behavioral-commitment verifier. In the passport installation map studied, 98.21% of passport dimensions lie in a nullspace and normalization adds a rescaling invariance, so one model yields many identities without establishing admission to a permissioned federation. A payload auditor over signed history detects six non-adaptive replay attacks across five held-out seeds, but also flags the honest client whose payload is copied, detects reduced training effort weakly, and is evaded by an attacker that evaluates its score. The results separate payload consistency, model identity and evidence that training occurred. Dataset Content: This repository contains the supplementary data and experimental assets to support and reproduce the results presented in the paper. Specifically, this dataset includes: Evaluation Results: The measured results supporting every table and figure in the manuscript, provided in JSON format. Archived Trajectories: Archived prototype trajectories that support the detector replay analyses. Model Checkpoints: Saved model checkpoints used for the passport-map measurements and inverse verifications.