ShaRP-FL: Decentralized Shapley-based Reward and Privacy Mechanism for Federated Learning

Federated learning (FL) enables clients to collaboratively train a model without sharing raw data, but practical systems must jointly address privacy, fair contribution measurement, and reliance on a trusted coordinator. Differential privacy (DP) protects training data by perturbing client updates, but conventional schemes typically allocate privacy without considering how strongly individual clients influence the global model. Shapley values provide a principled measure of marginal contribution~[1], but exact computation requires evaluating $2^{|K|}$ client coalitions and is typically centralized, creating a computational bottleneck and requiring clients to trust the server's validation data and scoring. Blockchain provides tamper-evident accounting, but does not by itself make contribution measurement decentralized or privacy-aware. We present \name, a unified FL protocol combining decentralized Shapley contribution measurement, contribution-adaptive DP, and blockchain-based authentication and accountability. Rather than evaluating all coalitions against a server-held validation set, \name\ distributes the $2^{|K|}$ coalition evaluations across $K$ clients, each using its own local validation data; neither training nor validation data leaves the client. The resulting Shapley scores determine contribution-proportional rewards and drive a tier-based adaptive DP mechanism. Higher-contribution clients receive smaller privacy budgets $\epsilon$ and correspondingly larger Gaussian noise multipliers $\sigma$, concentrating stronger protection on clients treated by our mechanism as more exposed to membership inference. We establish $(\epsilon,\delta)$-DP guarantees with respect to each client's training data for this adaptive allocation. A multi-signature smart contract authenticates participants through threshold voting, while model commitments, contribution scores, and rewards are recorded on-chain for tamper-evident accountability. We evaluate \name\ on MNIST, Fashion-MNIST, and CIFAR-10 under non-IID partitioning. The adaptive mechanism maintains competitive predictive utility while concentrating stronger privacy protection on high-contribution clients. Distributing coalition evaluation across $K$ clients achieves approximately $|K|\times$ wall-clock speedup, consistent with assigning approximately $2^{|K|}/|K|$ evaluations per client. A Likelihood Ratio Attack (LiRA)~[2] yields AUC $\approx 0.5$ across the evaluated datasets, indicating discrimination no better than random guessing under the evaluated attack setting. These results demonstrate that contribution measurement can jointly coordinate privacy and incentives while reducing dependence on centralized validation and accounting.