Single-Layer Trainable Neural Network for Secure Inference

Secure neural network inference provides privacy guarantees for both the client and the server, and is an integral approach in Machine Learning as a Service Setting (MLaaS). However, the multilayer structure in the neural network introduces frequent activation function calculations, which causes large overhead. Most of the prior secure inference systems focused on designing cryptographic protocols to improve computational efficiency, but high computing and communication overhead are still bottlenecks in practicality. In this work, we refocus on the potential of shallow neural networks and propose a model with only one trainable layer to reduce the required computation. Our main contributions are in three-fold: 1) introduce training-free weights and formally prove their contribution in the model expressivity; 2) design the Self Enhanced Module that is more suitable for shallow models as an alternative for the activation function; and 3) propose a linear layer with multiscale and normalization property, named Nested & Norm Conv. We conduct extensive experiments on visual datasets and the results demonstrate the proposed single-layer trainable model holds promise as a viable platform for secure inference in practical applications.