Federated learning (FL) trains a shared model without sending raw data, but some clients can be Byzantine and send harmful updates. Robust aggregation methods like Median and Krum can reduce poisoning damage, but they do not clearly show which client attacked. In this study, we propose RollupFL, an audit layer for FL that improves accountability under Byzantine attacks. RollupFL keeps aggregation and auditing separate, so it can work with FedAvg, Median, or Krum without changing how aggregation is computed. We study two audit designs: simple logging, which is fast, but assumes a trusted server, and blockchain-based audit, which gives stronger integrity and attribution, but adds more latency. We evaluate MNIST training for 20 rounds with 10%–30% Byzantine clients under sign-flip and model-replacement attacks. Results show that auditing does not meaningfully change accuracy, but it improves accountability. At 30% Byzantine, blockchain audit achieves higher attribution (0.95) and tamper detection (0.92) than logging (0.65 and 0.58). Logging adds small per-round latency, while blockchain adds larger latency mainly due to ledger writing.
Chowdhury et al. (Thu,) studied this question.