Privacy-centric digital surveillance through homomorphic encryption and deep learning
Johannes Unruh, Dorian Przetakiewicz, Oscar Hernán Ramírez-Agudelo, Michael Karl · 2025
We introduce LYNX (Layered privacY eNhancing eXchange), an open-source platform for privacy-preserving deep learning inference using homomorphic encryption (HE). LYNX enables end-to-end encrypted inference for neural networks by integrating Open Neural Network Exchange (ONNX) model support and TenSEAL-based secure computation. We demonstrate its practical application in surveillance scenarios like human detection, achieving real-time inference, all without exposing raw data. This paper presents the system architecture and implementation methodology, showcasing the feasibility of encrypted deep learning in privacy-critical applications.