Blockchain-Anchored Reputation for Robust Aggregation in Homomorphically Encrypted Federated Learning — Code and Data

Abstract

Blockchain-Anchored Reputation for Robust Aggregation in Homomorphically Encrypted Federated LearningCode, data splits, results and figures. Changes in version 1.1.0 Added code/tcga/, the preprocessing, audit, baseline and robustness scripts for the second corpus, and nibabel to requirements.txt. Added the second-corpus results: the label-flip and sign-flip result records, their summary tables and plots, and the centralised baseline log. Every second-corpus number in the paper was checked against these files. Replaced Figs. 1, 2, 3, 5, 6, 7 and 9 and Fig. A.1 with the versions in the paper. Fig. 3 no longer labels the reputation curve as encrypted, since the robustness runs aggregate in plaintext. Fig. 6(a) error bars now use the sample standard deviations of Table 14. Labels that overlapped data were moved. Added code/analysis/rebuild_figures_v2.py, which regenerates them. Added the graphical abstract and its source. Corrected two docstring lines in the five robustness scripts. They stated that an end-to-end encrypted run was performed separately, which it was not, and that leave-one-out scoring costs no additional privacy, whereas the key authority can difference two such aggregates. No code changed. Added a contract-guarantees section and a figure-to-paper table to the README, and a version note at the top of docs/threat_model.md.