Robust Representation Learning for Privacy-Preserving Machine Learning Using: A Multi-Objective Auto encoder Approach

Abstract

Privacy preservation has become a major challenge in modern machine learning because large-scale datasets often contain sensitive personal and organizational information. Conventional PrivacyPreserving Machine Learning (PPML) techniques, including Differential Privacy and Homomorphic Encryption, provide data protection but frequently compromise prediction accuracy, computational efficiency, or scalability. This project presents a robust representation learning framework based on a Multi-Objective Supervised Residual Autoencoder (SRAE) to achieve an effective balance between privacy preservation and model performance. The proposed framework transforms original data into compressed, discriminative feature representations that conceal sensitive information while retaining essential characteristics required for accurate learning. The model simultaneously optimizes multiple objective functions, including reconstruction loss, classification loss, centroid loss, and cosine similarity loss, to generate high-quality latent representations. These privacy-preserving features can be securely shared among multiple entities for model training and inference without exposing the original data. The framework is applicable to both unimodal and multimodal datasets and demonstrates strong performance on benchmark datasets such as MNIST, Fashion-MNIST, Retinal OCT, Leukemia, and TCGA Breast Cancer. Experimental evaluation shows that the proposed approach effectively protects data privacy while maintaining competitive classification accuracy and robust feature learning. This framework provides a practical, scalable, and secure solution for privacypreserving machine learning applications in healthcare, finance, and other data-sensitive domains.