FairWeight: Private Fairness-Aware Aggregation Using Model Interpretation Based Weightage in Federated Learning

Federated learning (FL) has been adopted as a promising solution for collaborative privacy-preserving machine learning. FL provides independence over the training data, but this can lead to statistical heterogeneity, which may result in issues such as bias and class imbalance. Thus, a global model, even with high accuracy, may be prone to discrimination and perform well only for certain sensitive attributes and classes. Global accuracy is not a perfect measure of the model's effectiveness; it should be reported in terms of balanced accuracy and fairness. Existing solutions attempt to mitigate bias using local or global debiasing techniques. However, local debiasing results in poor performance with increasing heterogeneity, and global debiasing is either designed only for achieving fairness, ignoring balanced accuracy, or costly in terms of finding an optimal solution. We present a novel model interpretation based bias mitigation aggregation scheme in FL, called FairWeight. It calculates the aggregation weightage by considering both fairness and loss for biased samples from the minority class. After optimizing the combined loss for prediction and fairness, the important weights responsible for bias and misclassification are calculated locally between clients. These weight indices are shared using secure aggregation to calculate the final aggregation weightage. Further, a secure aggregation scheme is devised that penalises biased weights across clients, rather than punishing the global model. FairWeight is evaluated on five real-world datasets and compared against five state-of-the-art fairness-aware aggregation schemes. The results provide empirical evidence that FairWeight achieves both balanced accuracy and fairness, with reduced computation and communication overheads.