Abstract
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.
Introduction
In modern wireless communications, such as 5G and beyond, Deep Learning (DL) has become a cornerstone for signal detection—the critical process of recovering transmitted bits from noisy and distorted signals. However, the efficacy of DL models is tied to the availability of massive datasets. Federated Learning (FL) offers a decentralized training paradigm where multiple users collaborate to train a global model without sharing raw data. Despite its advantages, FL faces a significant hurdle: data heterogeneity.
In a typical wireless network, each user experiences a unique Power Delay Profile (PDP) based on their specific environment. This variance causes local data distributions to differ significantly across users. During training, these differences pull local model updates in divergent directions, a phenomenon known as "client drift," which results in slow convergence and degraded accuracy.
The NCDSFL Methodology
To overcome these challenges, we introduce NCDSFL (Neural Collapse inspired Deep Supervised Federated Learning). This approach integrates two key architectural innovations to enforce feature alignment across diverse clients.
Neural Collapse as a Design Constraint
Neural Collapse (NC) is a phenomenon observed in over-parameterized models where, at the end of training, features collapse into a symmetric geometric structure, and classifier weights align perfectly with these features. Rather than allowing this to emerge naturally, NCDSFL mathematically derives the optimal NC solution for multi-binary classification and fixes the output layer weights to this state before training.
Deep Supervision and Feature Alignment
To ensure alignment throughout the network, we employ Deep Supervision (DS) by adding auxiliary shortcut classifiers to hidden layers. Like the final output layer, these auxiliary weights are fixed to an NC solution.
By fixing these "heads" across all users, the local feature extractors are forced to map heterogeneous wireless data into a shared, "collapsed" feature space. This effectively anchors the models and prevents drift during local training.
Experimental Setup
The study utilized simulated data based on industry-standard channel models:
- WINNER II Channel Model: Specifically the B1 non-line-of-sight (NLoS) scenario for urban environments.
- OFDM Parameters: 2.6 GHz carrier frequency with 20 MHz bandwidth.
- Data Diversity: 10 users were simulated, each with a unique PDP and 500 channel realizations for training.
Key Results and Discussion
The NCDSFL algorithm demonstrates significant improvements over standard Federated Learning (FedAvg) and Individual Learning (IL) baselines.
Faster Convergence and Accuracy
As shown in the performance analysis, NCDSFL reaches a stable Bit Error Rate (BER) significantly faster than standard FL. Specifically, NCDSFL converges in approximately 60 epochs, compared to the 150 epochs required by standard FL.
Robustness and Superiority
NCDSFL consistently achieved lower BER in heterogeneous scenarios. Furthermore, both FL and NCDSFL outperformed traditional Minimum Mean Squared Error (MMSE) signal detection. This superiority stems from the neural network's ability to learn the complex, implicit structures of OFDM channels that linear estimators often fail to capture. The system also exhibited robustness when tested against Rician channels, which were not part of the primary training set.
Conclusion
By transforming the observation of Neural Collapse into a proactive design constraint, NCDSFL effectively addresses the challenge of data heterogeneity in wireless federated learning. The use of fixed, mathematically optimal weights for deep supervision simplifies the optimization landscape and forces feature alignment across users. This results in a more robust, faster-converging signal detection model suitable for the complexities of future wireless environments.
Summary updated