Optimizing Privacy-Preserving Machine Learning: Comparative Insights from Encrypted Logistic Regression and SVM

This research presents a hybrid framework combining machine learning and cryptographic techniques for privacy-preserving spam email classification. Logistic Regression (LR) and Support Vector Machines (SVM) are employed, leveraging advanced preprocessing methods like tokenization, TF-IDF vectorization, and word embeddings for feature extraction. Homomorphic Encryption using the Paillier cryptosystem ensures computations on encrypted data, preserving privacy without compromising accuracy. LR demonstrates efficiency in training speed and interpretability, while SVM offers slightly better precision for imbalanced datasets. Encryption overhead challenges are addressed through modular arithmetic optimizations and sparse matrix representations. Evaluated on metrics such as accuracy, precision, recall, and F1-score, the framework balances robust classification with data security. This work highlights the feasibility of integrating cryptography into machine learning pipelines for secure and efficient spam detection, with potential applications in domains like healthcare and finance. Future work aims to enhance scalability, adopt transformer-based feature representation, and optimize real-time performance.