RePoSE: Retraining-Free Shapley Contribution Estimation in Decentralized Federated Learning

In Decentralized Federated Learning (DFL), participants collaborate without a central server. Each client trains on its local data and combines its model with those received from its neighbors. This collaboration has real costs (computation, energy, bandwidth), so participants need to be rewarded according to how much their local training improves the final models. Estimating this contribution is difficult in DFL, because no node observes all updates, local data differ across participants, and training may use differential privacy (DP), which adds noise to every update. Contributions are measured with the Shapley value, whose exact computation requires retraining the system for every subset of participants (coalition), a number that grows exponentially, so it is feasible only at small scale. This thesis proposes RePoSE (Reconstructed-coalition Post-hoc Shapley Estimation), a retraining-free estimator that uses the updates each participant records during training, without changing the training protocol. From these updates, each node reconstructs an approximation of its final model under every coalition and computes the Shapley values of the reconstructed game. RePoSE and the state-of-the-art decentralized estimator, used as the baseline, are measured against exact references from retraining every coalition, in simulations with eight participants on Fashion-MNIST and CIFAR-10, and under DP training on Fashion-MNIST. On Fashion-MNIST at thirty rounds RePoSE tracks the ranking of the exact reference in every condition, while the baseline orders the participants against it in three quarters of them. RePoSE allocates 80% of the reward budget as the reference does, against 40% for the baseline. On CIFAR-10 the advantage is smaller, 72% against 59%. Under DP training RePoSE stays as faithful as without privacy and leads the baseline in every condition, with 84% against 58%. Its added traffic is half of the baseline's after ten rounds and one sixth after thirty.