A new study published in Applied Intelligence proposes a solution called Hybrid Fair, a five-stage aggregation framework designed to handle update robustness, per-client metric fairness and on-chain storage overhead within a single pipeline. The work was carried out by Ripunjay Sharma, Shabd Sahay, Yatendra Sahu, Gagan Vishwakarma and Arpita Bhargava, and appeared online on 27 October 2026.

The framework begins by screening out anomalous updates using a combination of Z-Score and Mahalanobis distance filtering, a step intended to keep unreliable or corrupted client contributions from distorting the shared model. It then assigns adaptive weights to each client based on a round-dependent blend of accuracy and F1 score, with logarithmic dampening applied according to dataset size so that clients with very large datasets do not dominate the aggregation.

A distinctive feature of the method is its fairness multiplier, which penalises clients whose accuracy and F1 score are poorly balanced. This is aimed at sustaining balanced precision, recall and client-level fairness simultaneously, rather than allowing one metric to improve at the expense of another when local data distributions are heterogeneous.

To keep the blockchain component affordable, the system records only a 32-byte SHA-256 hash of the aggregated model on a Solidity smart contract each round, instead of storing full model parameters on-chain. According to the authors, this hash-anchoring design reduces gas consumption by 97.33 percent relative to naive full-parameter storage, and the saving is architecture-agnostic. An extended comparison against 8-bit parameter quantization and Merkle-tree batch verification is also reported in the paper.

The experimental evaluation used two publicly available fraud benchmarks, the Credit Card Fraud dataset and PaySim Mobile Money, across four model architectures, seven baseline aggregation strategies and ten independent random seeds. Hybrid Fair achieved the highest F1 score in six of eight model–dataset combinations.

The reported gains were strongest on the Credit Card Fraud benchmark, where the framework showed statistically significant improvements over all seven baselines for MLP, with a 3.7 percent gain at p ≤ 0.001, and for XGBoost, with a 4.1 percent gain at p ≤ 0.001. On PaySim, the improvements were more selective: significant advantages over three of the seven baselines were observed for Random Forest, with the best result a p value of 0.010 against FLAME, and for XGBoost, with the best result a p value of 0.005 against FLTrust.

The authors also note a larger but statistically less certain margin on PaySim, where Hybrid Fair’s own advantage over FedAvg reached 9.3 percent for Random Forest and 45.0 percent for XGBoost, although the FedAvg comparison itself did not individually reach significance at this seed count. That caveat matters, because it means the headline margins on PaySim should be read as descriptive results rather than confirmed effects, and further replication with more seeds would be needed to establish them firmly.

Overall, the study presents a case that aggregation, fairness and blockchain cost need not be traded off against one another in federated systems. By combining anomaly filtering, metric-balanced weighting and hash-only on-chain anchoring, the framework offers a template for deployments where heterogeneous client data and limited on-chain budgets coexist, though its evidence base remains confined to two fraud benchmarks and simulated federated settings.

Subject of Research: A federated learning aggregation framework that combines anomaly filtering, fairness weighting and hash-based blockchain anchoring, evaluated on fraud detection benchmarks.

Article Title: Adaptive multi-metric aggregation for cost-efficient and fair blockchain-federated learning

Article References: Sharma, R., Sahay, S., Sahu, Y., Vishwakarma, G., & Bhargava, A. (2026). Adaptive multi-metric aggregation for cost-efficient and fair blockchain-federated learning. Applied Intelligence, 56(16), 521. https://doi.org/10.1007/s10489-026-07561-y

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07561-y

Keywords: federated learning, blockchain, aggregation, fairness, machine learning, F1 score, gas consumption, smart contracts, fraud detection, PaySim, credit card fraud, XGBoost