Dataset of round-level federated learning and layer-2 blockchain overhead from ten raspberry Pi edge clients

Talgar Bayan, Bibarys Mukhambetiyar, Askar Boranbayev, Adnan Yazıcı

Data in Brief, 2026

Full text Open access: Gold Metadata (CSL-JSON)
DOI 10.1016/j.dib.2026.113111
Data confidence 85 high
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

This data article describes round-level records from 40 federated learning sessions run on a physical testbed of ten Raspberry Pi 4 Model B devices acting as edge clients, with an Ubuntu workstation as the aggregation server. Each session trains a one-dimensional Squeeze-and-Excitation ResNet classifier over ten aggregation rounds using the Flower framework and commits every round to a smart contract on Base mainnet, a public Ethereum-compatible Layer-2 chain, while pinning the model and metric artifacts to IPFS through Pinata. The sessions cover two human activity recognition benchmarks, MHEALTH and UCI-HAR, and two middleware configurations: a baseline that performs blockchain and IPFS operations synchronously with per-round gas estimation, and an optimised variant that overlaps IPFS uploads with local training and reuses cached gas parameters, giving 20 sessions per benchmark. For every round the dataset records server-side global metrics (accuracy, F1, AUC, and the confusion matrix), per-device training telemetry (loss, training time, CPU, memory, and temperature), per-client evaluation records, and the IPFS content identifiers and transaction hashes that link each round to Base mainnet. For every session, millisecond-resolution logs record the latency of each blockchain and storage operation. The repository contains 12 CSV files, two archives of raw session directories, and a data dictionary. Researchers can reuse the convergence traces to benchmark edge federated learning, the device telemetry to characterise resource use on constrained hardware, and the latency distributions to parameterise simulations of blockchain-enabled federated learning without physical deployments or transaction fees. All 1240 blockchain transactions remain publicly verifiable on the Base mainnet explorer.