Speaker:Pinglan Liu

Securing Outsourced Computation Based on Game Theory, HE and TEE

With the popularity and prosperity of cloud services, outsourcing computation to cloud servers has become an irreversible trend that can not only reduce the computation burden of clients, but also fully utilize the resources of and make profits for servers. However, the cloud servers may not be fully trustworthy. For example, due to external attacks or internal malice, cloud servers may steal sensitive information involved in the outsourced computation, or may not execute the computation honestly but instead fake the results. Hence, it is desired to retain information confidentiality and computation integrity. In this dissertation, we will focus on constructing effective and efficient solutions for preserving confidentiality and assuring integrity in computation outsourcing. We are particularly interested in exploring the solutions based on the integration of homomorphic encryption (HE), trusted execution environment (TEE) and game theory.

Our research includes two parts. In the first part, we propose game-theoretic schemes to protect the integrity of outsourced generic computation, which includes: (1) a scheme with trusted third party (TTP); (2) a scheme with blockchain instead of TTP; (3) algorithms for optimizing fund allocation in game theory based computation outsourcing. In the second part, we target at protecting the confidentiality of data and model for the outsourced inference and training of neural network models, based on the integration of leveled-HE (LHE) and TEE. We start with a new framework that enables collaboration among parties that do not trust each other. We also propose a generic and efficient LHE-based inference scheme, along with optimizations, as an important performance-determining component of the framework. Then we extend our work with a model-refining scheme that integrates LHE and TEE. The core of the scheme is the application of LHE for protecting the confidentiality of training data and model in the backward propagation process of model training, and the application of TEE for periodically reducing noises in ciphertexts. We have conducted comprehensive quantification and verification of the security of our proposed schemes, and detailed evaluations of their performance efficiency.

Committee: Wensheng Zhang (major professor), Wallapak Tavanapong, Yong Guan, Ying Cai, and Jia Liu.

Join on Zoom: https://iastate.zoom.us/j/92977823189