FedCBA: Contribution-based Aggregation in Decentralized Federated Learning
Authors
Eberhardt, Rune Iversen ; Pedersen, Lucas Lybek Højlund ; Jensen, Caspar Emil
Term
4. term
Education
Publication year
2026
Submitted on
2026-06-11
Abstract
This thesis introduces Federated Contribution-based Aggregation (FedCBA), a new aggregation framework for decentralized federated learning. Federated learning is a way to train a shared global model without collecting all data in one place: each participant trains locally and only sends model updates, not raw data. FedCBA focuses on measuring how much each participant truly contributes to the global model and uses this to weight their updates and to encourage honest behavior. Instead of relying on self-reported contribution metrics, FedCBA uses peer-based evaluation: participants evaluate each other’s model updates. This allows the system to compute a contribution score for every participant without depending on potentially biased self-reports. We design and test several aggregation strategies that weight model updates according to these contribution scores. These strategies are then combined into an adaptive switching mechanism that can automatically change between methods and adjust their merge weights, depending on how the global model is progressing during training. FedCBA is evaluated in settings where some participants behave adversarially, including malicious actors and freeriders who try to benefit from the global model without contributing fairly. These adversarial participants carry out different types of attacks on the training process. To assess how robust and effective FedCBA is, we run a series of experiments on the widely used image datasets MNIST and CIFAR-10. Two of the experiments directly evaluate the proposed aggregation strategies. In the first experiment, we disable FedCBA’s incentive mechanisms to isolate and assess the aggregation strategies on their own. In the second experiment, we deploy the full FedCBA system and evaluate how the aggregation strategies perform when combined with the incentive mechanisms. Across these experiments, the empirical results, Area Under the Curve analyses, and Wilcoxon signed-rank tests show that our contribution-weighted strategies achieve statistically significant improvements in global model accuracy and faster loss reduction compared to FedAvg, which serves as the baseline throughout.
Dette speciale præsenterer Federated Contribution-based Aggregation (FedCBA), en ny metode til at samle (aggregere) modeller i decentraliseret federated learning. Federated learning er en måde at træne en fælles, global model på, uden at deltagerne deler deres rå data. I stedet træner hver deltager lokalt og sender kun modelopdateringer videre. FedCBA fokuserer på at vurdere, hvor meget hver deltager faktisk bidrager til den fælles model, og bruger denne vurdering til både at vægte deres bidrag og til at motivere ærlig opførsel. I stedet for at deltagerne selv rapporterer, hvor meget de har bidraget, anvender FedCBA såkaldt peer-baseret evaluering: Deltagerne vurderer hinandens modelopdateringer. På den måde kan systemet beregne en bidragsscore for hver deltager uden at være afhængig af selvrapporterede tal, som kan være fordrejede. Vi designer og afprøver flere forskellige aggregationsmetoder, hvor deltagernes opdateringer vægtes efter deres beregnede bidrag. Disse metoder kombineres i et adaptivt system, som automatisk kan skifte mellem strategier og justere vægtningen, afhængigt af hvordan den globale model udvikler sig over tid. FedCBA testes i scenarier, hvor nogle deltagere opfører sig ondsindet eller forsøger at freeride – det vil sige drage fordel af den fælles model uden selv at bidrage ordentligt. Disse ”adversariske” deltagere udfører forskellige typer angreb på træningsprocessen. For at undersøge, hvor robust og effektiv FedCBA er, gennemfører vi en række eksperimenter på de velkendte billeddatasæt MNIST og CIFAR-10. To af eksperimenterne undersøger direkte, hvordan de foreslåede aggregationsstrategier klarer sig. I det første eksperiment slår vi FedCBA’s incitamentsmekanismer fra for at isolere og teste selve aggregationsstrategierne. I det andet eksperiment anvendes hele FedCBA-systemet, hvor aggregationsstrategierne evalueres sammen med incitamentsmekanismerne. I disse eksperimenter viser både de observerede resultater, Area Under the Curve-analyser og Wilcoxon signed-rank tests, at vores bidragsvægtede strategier giver statistisk signifikante forbedringer i den globale models nøjagtighed og hurtigere reduktion af tab (loss) sammenlignet med FedAvg, som bruges som baseline gennem alle forsøg.
[This abstract has been rewritten with the help of AI based on the project's original abstract]
Keywords