Federated Learning for Arabic Sentiment Analysis

Our paper evaluates lightweight Federated Learning (FL) models for privacy-preserving Arabic sentiment analysis. We propose a CPU-efficient FL framework using classical models (SVM, Random Forest, Logistic Regression, etc.) as a practical alternative to resource-intensive transformers. tested on Arabic tweet datasets (ArSAS, ASTC, SS2030), Federated SVM achieved top accuracy (e.g., ~90% on ArSAS), while Logistic Regression offered an optimal balance of high performance and minimal execution time. Results confirm FL provides comparable accuracy to centralized training with no significant performance trade-off, while ensuring data privacy, low communication costs, and scalability, making it ideal for edge devices.