Privacy-Preserving Automated Deep Learning for Secure Inference Service

Summary

The paper presents 2PCAutoDL, a privacy-preserving automated deep learning scheme that searches DNN architectures on encrypted data using two non-colluding cloud servers. Its optimized secure protocols report 1.34x to 2.05x speedups for linear layers and 1.33x to 45x for non-linear layers, plus a Softmax approximation with slightly higher accuracy. Under the semi-honest adversary model, searched models improve inference accuracy by 0.6% on MNIST and 0.5% on CIFAR-10 over state-of-the-art PrivAutoDL.