Delphi: A Cryptographic Inference Service for Neural Networks

By Pratyush Mishra, Ryan T. Lehmkuhl, Akshayaram Srinivasan, Wenting Zheng, Raluca Ada Popa. 2020

605 citations recorded by the source corpus.

Many companies provide neural network prediction services to users for a wide range of applications. However, current prediction systems compromise one party’s privacy: either the user has to send sensitive inputs to the service provider for classiﬁcation, or the service provider must store its proprietary neural networks on the user’s device. The former harms the personal privacy of the user, while the latter reveals the service provider’s proprietary model. We design, implement, and evaluate D ELPHI , a secure prediction system that allows two parties to execute neural network inference without revealing either party’s data. D ELPHI approaches the problem by simultaneously co-designing cryptography and machine learning. We ﬁrst design a hybrid cryptographic protocol that improves upon the communication and computation costs over prior work. Second, we develop a planner that automatically generates neural network architecture conﬁgurations that navigate the performance-accuracy trade-offs of our hybrid protocol. Together, these techniques allow us to achieve a 22 × improvement in online prediction latency compared to the state-of-the-art prior work.

Subfields: Artificial Intelligence, Computer Science Applications, Sociology and Political Science

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Science Applications

In Sociology and Political Science

Citations in the corpus

Falcon: Honest-Majority Maliciously Secure Framework for Private Deep Learning (2020, neutral): “Delphi [10] # # # # # # # # # # G #”
CrypTFlow2: Practical 2-Party Secure Inference (2020, background): “Appen-dix G provides a brief summary of these benchmarks.”
A Principled Approach to Learning Stochastic Representations for Privacy in Deep Neural Inference (2020, neutral): “DELPHI [9] 2020 ResNet-32 CIFAR-100 3.”
CrypTFlow: Secure TensorFlow Inference (2019, methodology): “Although the use of fixed-point for secure computations is well-known [68], prior works on secure inference have addressed the float-to-fixed problem by either generating a fixed-point model by hand ([62], [55], [49], [60], [79], [58], [72]), or by using non-standard training algorithms that output ”
Cheetah: Optimizing and Accelerating Homomorphic Encryption for Private Inference (2020, background): “However, MPC performance is limited by communication bottlenecks [37], [42], [51], which require consideration of network-protocol and technology levels, or redesigning neural architectures [28], [41].”
CrypTen: Secure Multi-Party Computation Meets Machine Learning (2021, methodology): “These systems include Gazelle [38], Chameleon [58], EzPC [15], MiniONN [45], SecureML [49], PySyft [60], and Delphi [47].”
Backward Feature Correction: How Deep Learning Performs Deep Learning (2020, background): “Note 2: Our obtained accuracies using quadratic networks may be of independent interests: networks with quadratic activations have certain practical advantage especially in cryptographic applications [61].”