Vyshnavi Mattewada,Mrs. J. Shilpa 03 06 Protecting sensitive information from unauthorized exposure has become a critical challenge in modern cybersecurity environments Conventional data leakage detection mechanisms often require access to unencrypted content creating potential privacy and security risks during the inspection process This paper proposes a privacyaware data leakage detection framework that integrates Fully Homomorphic Encryption FHE with advanced deep learning techniques to enable secure analysis of protected data without revealing its original content The proposed model named HEDLDNet utilizes a transformerinspired neural architecture specifically adapted to process encrypted information while maintaining strong confidentiality guarantees By performing detection directly on encrypted data streams the framework eliminates the need for decryption during analysis and significantly reduces the risk of sensitive information exposure Extensive experimental evaluation demonstrates that the proposed approach achieves a detection accuracy of 943 while maintaining practical computational performance Performance comparisons indicate that the framework lowers processing overhead by approximately 73 relative to existing homomorphic encryptionbased detection methods In addition the system supports near realtime monitoring with response delays below 250 milliseconds making it suitable for deployment in enterprisescale environments The results highlight the effectiveness of combining privacypreserving cryptographic techniques with intelligent learning models to build secure and efficient data leakage detection solutions This research contributes toward the development of nextgeneration cybersecurity systems that balance strong privacy protection with accurate threat identification Copyright (c) Journal of Science Engineering Technology and Management Science. All rights reserved 2026