Federated Learning (FL) ensures privacy by training models on distributed devices without sharing raw data. However, privacy concerns arise as adversaries can still extract information from model parameters. This is crucial for security-critical applications like speech emotion recognition (SER). Existing privacy mechanisms impact system performance, leading to slower responses and reduced accuracy. Balancing privacy and performance is essential, and this research aims to propose privacy-preserving methods for FL in SER while addressing performance challenges. This research focuses on finding the right combination of methods to safeguard sensitive information without compromising system efficiency.