Gecko: Fast Private Inference via Secure Public Encoder Offloading

Gecko is a private inference framework that keeps a small predictor protected while offloading a frozen public encoder outside the trust boundary.

It combines hierarchical features, Fastfood compression, and private feature gating to make encrypted inference faster and lighter.

The system reports 0.4–2.2 second inference times with up to 10.8 MB communication on image and audio tasks.

The paper argues that naive encoder offloading can create extraction risks, and Gecko is designed to preserve efficiency without weakening security.