Cloud-based machine learning faces a tension: users want their inputs kept private, while service providers want to protect their models. Fully homomorphic encryption (FHE) offers a way to compute on encrypted data, but it has historically been too expensive for practical inference. A new arXiv paper, "Practical and Private Hybrid ML Inference with Fully Homomorphic Encryption," explores a hybrid strategy that mixes FHE with plaintext operations.
By splitting the computation, the approach aims to keep sensitive user data encrypted while limiting costly FHE operations to the parts that need them. This also helps preserve the confidentiality of the server-side model, since the full model is not exposed to the client. The abstract frames the work around the two core requirements of modern cloud services: protecting user data and keeping the model secret.
The paper's contribution is in making FHE-based inference more feasible rather than treating it as an all-or-nothing solution. The abstract does not provide performance figures, but the emphasis on a hybrid design suggests a practical middle ground between full encryption and conventional plaintext inference.