CorrFL: Correlation-Based Neural Network Architecture for Unavailability Concerns in a Heterogeneous IoT Environment

The Federated Learning (FL) paradigm faces several challenges that limit its application in real-world environments. These include the local models' architecture heterogeneity and unavailability of distributed Internet Things (IoT) nodes due to connectivity problems. factors posit question "how can available models fill training gap unavailable models?". This is referred as "Oblique Learning" problem. problem encountered studied environment includes IoT responsible for predicting CO2 concentrations. paper proposes Correlation-based FL (CorrFL) approach influenced by representational learning field address this CorrFL projects various model weights a common latent space heterogeneity. Its loss function minimizes reconstruction when are absent maximizes correlation between generated models. latter factor critical because intersection feature spaces devices. evaluated on realistic use case, involving one device heightened activity levels reflect occupancy. from ones trained new compared against different cases, benchmark model. evaluation criteria combine mean absolute error (MAE) predictions impact amount exchanged data prediction performance improvement. Through comprehensive experimental procedure, outperformed every criterion.