Collaborative fairness, as a fundamental requirement in federated learning, incentivizes active client participation in collaborative model training through equitable reward distribution. However, many methods ignore data distribution traits when evaluating client data quality, leading to reward allocations that mismatch with actual contributions under heterogeneous data conditions. This mismatch worsens the impact of low-quality data on global model optimization and lowers long-term client engagement. To tackle these issues, we present FedDAR, a framework balancing fairness and performance effectively. FedDAR pairs an adaptive reputation network with the Multifactor Evaluation Score (MES) mechanism to dynamically assess client contributions across four aspects: past reputation, data size, diversity, and model performance. Additionally, it incorporates a dynamic gradient adjustment module based on Jensen-Shannon divergence to ensure alignment between contributions and rewards. Experimental results on the MNIST, CIFAR-10, and EMNIST datasets demonstrate that FedDAR outperforms baseline methods in collaborative fairness while achieving prediction accuracy comparable to state-of-the-art approaches and consistently surpassing the standalone framework. This dual advantage underscores FedDAR’s robust adaptability to heterogeneous data dynamics.