Most existing Federated Learning (FL) methods excessively rely on features associated with sensitive attributes, prioritizing global accuracy while neglecting issues like class imbalance, which can undermine both group and individual fairness. Additionally, these methods often attribute client contributions solely to data volume, overlooking other critical factors, such as data quality and per-client performance. To address these issues, this paper proposes a fairness-aware federated learning based on feature attention and contribution calibration (FAFLC). FAFLC extracts feature weights using a feature attention mechanism. It calculates the output of each feature through a multilayer perceptron, maps the results to an exponential function, and normalizes them to ensure consistency across features. These weights are incorporated into a weighted loss function, which reduces reliance on sensitive attributes and promotes more balanced usage of features. Then the client contribution calibration method evaluates factors including local data volume, balanced accuracy, and fairness metrics during local model training to ensure a more accurate assessment of each client’s contribution. The weights of these dimensions are dynamically adjusted based on changes in client performance during training, thereby optimizing both the global model quality and fairness. Experimental results show that FAFLC improves balanced accuracy by an average of 6.7% across four real-world datasets. Additionally, FAFLC’s demographic parity scores reduced by an average of 64.24%, approaching zero.