FedAdaptSparse: A Sparse Federated Learning Approach to Reduce Communication Overhead in Emotion Detection Systems

Marron G.A.C.; Mamani-Aliaga A.H.; Bustincio R.W.C.; Junior W.V.L.; Bittencourt L.F. · Proceedings - 2026 22nd International Conference on Distributed Computing in Smart Systems and the Internet of Things, DCOSS-IoT 2026 (2026) · DOI: 10.1109/dcoss-iot69657.2026.00171

Resumen

Federated Learning (FL) has emerged as a key paradigm for privacy-preserving emotion detection in sensitive mental health datasets. However, the high communication overhead of modern deep learning architectures, such as hybrid CNN-LSTM models, poses a significant scalability challenge in communication overhead-constrained environments. This paper proposes FedAdaptSparse, a Sparse Federated Learning approach based on Static Sparsity Federated Learning (SSFL) principles. The core logic relies on identifying a global saliency mask during an initial "Round 0"to select a critical subset of parameters for training, effectively reducing the transmission payload. Using a dataset of 35,000 social media posts labeled across 6 mental health categories (Normal, Depression, Suicidal, Anxiety, Stress, Bipolar), we simulate a non-IID federated scenario with 10 clients. Experimental results over 30 communication rounds demonstrate that FedAdaptSparse achieves a communication reduction of 90.13% compared to a dense baseline, while maintaining a competitive accuracy of 40.95% (0.8% gap). This reduction is particularly relevant for IoMT-enabled public hospital edge deployments, where multiple hospital nodes share limited communication overhead for collaborative model training on distributed patient data. Our findings validate FedAdaptSparse as a viable strategy for efficient emotion detection in communication overhead-constrained urban computing scenarios.

Instituciones: Universidad Nacional de San Agustín de Arequipa