FL-VetTrans: Privacy-Preserving Translation of Animal Vocalization for Clinical Diagnosis via Federated Learning

With the growing integration of artificial intelligence in veterinary healthcare, automated analysis of animal vocalizations offers a promising non-invasive approach for early disease detection. However, traditional centralized learning frameworks raise privacy concerns and struggle with non-IID data distributions across distributed clinical environments. In this study, we propose FL-VetTrans, a privacy-preserving federated learning framework for disease classification using domestic animal vocal signals. The system leverages log Mel-Spectrogram representations of raw audio recordings and employs a convolutional neural network (CNN) for end-to-end acoustic feature learning. Training is performed locally at distributed edge nodes without sharing raw data. To address statistical heterogeneity across clients, FedProx regularization is incorporated to stabilize model convergence under non-IID conditions. Furthermore, differential privacy is implemented using Opacus-based DP-SGD, ensuring formal privacy guarantees through gradient clipping and noise injection. Experimental evaluation on healthy and unhealthy poultry vocal datasets demonstrates that the proposed federated framework achieves 93\% global classification accuracy, along with strong precision, recall, F1-score, and ROC-AUC performance. These results highlight the effectiveness of privacy-preserving federated acoustic modeling for veterinary disease diagnosis.