Sensitive data is typically protected while stored or transmitted, but must be decrypted for processing in cloud environments. This creates a critical security gap, particularly for applications such as machine learning on medical or financial data. Fully Homomorphic Encryption (FHE) aims to close this gap by enabling computation directly on encrypted data, without exposing it at any stage.

This talk provides a practical introduction to FHE and explores its current feasibility in real-world scenarios. Using the Concrete-ML framework, an end-to-end example of encrypted machine learning inference is demonstrated, where a model is deployed in a cloud-like environment and processes encrypted input without access to the underlying data.

The session focuses on hands-on insights where key aspects such as performance overhead, accuracy, and scalability are examined based on experimental results. The findings show that while FHE enables strong data confidentiality and works well for certain classes of models, it introduces significant computational costs and practical limitations.

The talk concludes with a discussion of where FHE can already be applied today, where it remains impractical, and what developments are needed to make privacy-preserving cloud computing more widely usable.