S&P 2022

Poster: Secure SqueezeNet inference in 4 minutes

Try demo


Privacy-preserving machine learning (PPML) solutions often use multi-party computation or client-assisted homomorphic encryption (HE) techniques, which require substantial communication overheads. In contrast, non-interactive solutions are considered slow and are practical for small neural networks or with limited security guarantees. In this work, we show for the first time that it is possible to evaluate a large HE-friendly SqueezeNet model on large images in a non-interactive setting using HE, with 128-bits security parameters in only 4 minutes when running on a GPU and 6 minutes when running on a CPU.