Neural Collapse Based Deep Supervised Federated Learning for Signal Detection in OFDM Systems

Future wireless networks are expected to be AI-empowered, making their performance highly dependent on the quality of training datasets. However, physical-layer entities often observe only partial wireless environments characterized by different power delay profiles. Federated learning is capable of addressing this limited observability, but often struggles with data heterogeneity. To tackle this challenge, we propose a neural collapse (NC) inspired deep supervised federated learning (NCDSFL) algorithm. Specifically, we first define an NC solution for the multi-binary classification problem and establish its optimality by revealing its connection to the global optimum. We then incorporate the deep supervision technique into deep neural networks and fix the weights at both the output layer and an auxiliary hidden layer using the derived NC solutions. When this strategy is applied in a federated learning setting, it encourages different clients to produce similar hidden features, thereby enabling the proposed algorithm to effectively address data heterogeneity and achieve fast convergence. Simulations for signal detection in OFDM systems confirm the NC phenomenon and demonstrate that NCDSFL outperforms several baselines in terms of convergence speed and accuracy.