GPAFed: Gradient Projection-Guided Adaptive Aggregation Strategy for Federated Learning Under non-IID Data

Under non-independent and identically distributed (non-IID) data conditions, static weighted aggregation strategies in Federated Learning (FL) often fail to reflect each client's effective contribution to the global objective. This limitation may induce gradient conflicts, leading to unstable convergence and model performance degradation. To address this issue, we propose a Gradient Projection-guided Adaptive aggregation strategy for Federated Learning (GPAFed). We derive a novel convergence bound that explicitly incorporates the projection of local gradients onto the global gradient, revealing its quantitative influence on global loss reduction in FL. Building on this insight, we formulate a projection-based client contribution metric and design an adaptive aggregation weighting strategy inspired by the softmax mechanism. This approach strengthens the influence of high-contribution clients while attenuating the adverse effects of low-quality updates. Finally, we provide a theoretical guarantee that the proposed strategy achieves a tighter convergence bound. Extensive experiments under diverse non-IID conditions demonstrate that GPAFed consistently outperforms strong baselines in terms of model accuracy, convergence speed, and robustness.