Federated privacy-preserving facial paralysis severity classification using hybrid deep learning and pufferfish optimization

Abstract Facial paralysis (FP) severity assessment requires accurate identification of subtle facial asymmetry while maintaining the confidentiality of sensitive patient information. Conventional centralized deep learning approaches require patient images to be transferred to a central repository, which can increase privacy risks and limit collaborative learning across healthcare institutions. To address these challenges, this study proposes a Federated Learning-based Secure and Privacy-Preserving Hybrid Deep Learning with Pufferfish Optimization (FLSDPP-HDLPO) framework for automated FP severity classification. The proposed framework first performs image preprocessing and augmentation to improve facial-image quality and class representation. A hybrid feature extraction strategy integrating Capsule Network (CapsNet), ResNet50, and LeViT is then employed to learn complementary spatial, hierarchical, and attention-based facial representations. The extracted features are classified using a CNN-BiGRU architecture, while the Pufferfish Optimization Algorithm (POA) is employed to optimize the relevant model parameters. To preserve data confidentiality during distributed learning, homomorphic encryption is incorporated into the federated learning process so that sensitive model information can be securely exchanged without exposing the original patient data. The framework is evaluated using combined MEEI and YFP facial paralysis datasets containing six severity classes. Experimental evaluation under different training and testing configurations demonstrates that the proposed FLSDPP-HDLPO framework achieves a maximum classification accuracy of 97.56%, with precision, sensitivity, specificity, and F1-score of 92.48%, 92.53%, 98.54%, and 92.50%, respectively. Repeated experiments further produced an accuracy of 97.56% ± 0.42%, demonstrating consistent performance. The proposed framework therefore provides an accurate and privacy-preserving approach for distributed FP severity assessment. Future work will focus on larger multi-center datasets, stronger robustness evaluation, computational optimization, and real-world clinical validation.