Data leakage poses significant threats to organizational security, with traditional detection methods requiringaccess to sensitive information in plaintext. This paper presents a novel framework for privacy-preserving data leakagedetection that combines fully homomorphic encryption (FHE) with deep learning architectures. Our approach enablesorganizations to detect sensitive data exposure without decrypting the monitored content, addressing critical privacyconcerns. We introduce HE-DLDNet, a modified transformer-based neural network that operates directly on encrypted datastreams, achieving 94.3% detection accuracy while maintaining computational efficiency. Experimental results onbenchmark datasets demonstrate that our method reduces computational overhead by 73% compared to existinghomomorphic approaches while preserving privacy guarantees. The framework supports real-time detection with latencyunder 250ms, making it suitable for production deployment. This work bridges the gap between strong privacy preservationand practical data leakage detection systems.
P. et al. (Mon,) studied this question.