Assessing the Impact of Fully Homomorphic Encryption on Machine Learning Performance: A Case Study in Titanic Survival Prediction

Privacy in machine learning is essential as personal data usage grows. Fully Homomorphic Encryption (FHE) enables encrypted computations but is computationally costly. Concrete-ML integrates FHE into machine learning while maintaining functionality. This study examines its impact on XGBoost, Random Forest, and Decision Tree models using the Titanic dataset. Results show Decision Tree maintains high accuracy (79.43%–81.10%) with better efficiency, while XGBoost and Random Forest have longer inference times. Despite strong privacy protection, FHE's high computational cost hinders practical use. Future research should explore GPU/TPU acceleration and optimized encryption protocols for improved efficiency.