Over the past decade, computer vision has become foundational across applications such as facial recognition, medical diagnostics, surveillance, and social media. As these systems grow more advanced and proprietary, concerns around privacy and security intensify. For example, when a user uploads a facial image to a cloud-based recognition service, they want to protect their biometric identity, while the service provider must safeguard its proprietary model and private data. This creates complex but critical challenges.

Data encryption is widely used to address privacy concerns, but conventional schemes require data decryption for processing, which introduces security risks. Fully Homomorphic Encryption (FHE) offers a transformative alternative by enabling computation directly on encrypted data, keeping sensitive information hidden throughout. It also provides post-quantum security, making it resilient to future quantum threats. This is especially valuable in privacy-critical domains such as healthcare, where regulations like HIPAA must be met, or in partially trusted cloud environments.

Integrating computer vision (CV) with FHE is challenging due to the complexity of CV models and FHE’s limited support for arithmetic operations like addition and multiplication. These operations are insufficient for nonlinear functions common in CV models, such as ReLU, batch normalization, and pooling, which FHE does not natively support. Moreover, encrypted computation is computationally intensive and memory-demanding, leading to high latency and large memory usage. Designing effective CV systems under FHE constraints requires balancing accuracy, efficiency, and resource use, while also creating architectures compatible with FHE.

Despite these challenges, substantial progress has been made in developing practical CV applications with FHE. Notable examples include private image retrieval using encrypted embeddings, encrypted image classification, and fully encrypted face recognition systems. These advances stem from both improvements in FHE schemes and thoughtful redesigns of traditional CV models to fit FHE constraints. Their success highlights the growing feasibility and importance of secure, privacy-preserving computer vision.

This tutorial is designed to offer a detailed and accessible introduction to the field of encrypted computer vision using FHE and will cover the following topics.

The tutorial will also highlight emerging research at the intersection of computer vision, machine learning, and homomorphic encryption. By the end, participants will gain the knowledge and tools to explore and develop secure, privacy-aware vision applications.