CrowdFORGE: Federated Learning for Optimizing Crowdsourced Contributions in Centralized Global Models

Abstract

The increasing use of distributed, crowdsourced data in federated learning introduces major challenges related to non-IID data distributions, user privacy, and limited computational capacity on resource-constrained devices. Traditional federated learning approaches often sacrifice either model accuracy or privacy preservation to maintain system efficiency. To address these limitations, we present CrowdForge, a federated learning framework designed to maximize global model accuracy through intelligent aggregation of heterogeneous crowdsourced inputs without compromising privacy or device performance.

CrowdForge integrates multiple adaptive aggregation strategies—including FedAvg, FedProx, and AdaptiveFedAvg—to dynamically select the most suitable method based on client characteristics and data quality. Privacy is ensured through the incorporation of Differential Privacy and Secure Multi-Party Computation, enabling secure model updates while minimizing the loss of utility. Furthermore, CrowdForge employs quality-aware filtering and lightweight feature engineering to ensure that only high-value client contributions influence the global model, thereby improving robustness in highly diverse environments.

Experimental evaluations demonstrate that CrowdForge improves model accuracy by an average of 8.2% across diverse non-IID datasets when compared to conventional federated learning baselines. The framework additionally reduces communication overhead by 35%, significantly enhancing training efficiency for low-resource devices that participate intermittently or possess limited bandwidth.

Benchmark results further confirm that CrowdForge sustains 90% model performance on CIFAR-10 under non-IID partitions, outperforming standard federated learning methods by 6% while upholding strong privacy guarantees. CrowdForge also effectively handles device-level resource scarcity by reducing per-device computation by 30% through adaptive learning strategies and selective model update mechanisms.

Overall, these findings establish CrowdForge as a scalable, efficient, and privacy-preserving federated learning system capable of achieving high global model accuracy in challenging real-world environments characterized by heterogeneous data and limited computational resources.