With the rapid development of machine learning, MLaaS has infiltrated into many fields such as image recognition, natural language processing, medical diagnosis, and so on. However, in MLaaS, data interaction between users and service providers is inevitable, and both users' private data and servers' model parameters are at risk of privacy disclosure. In order to solve this problem, homomorphic encryption is an extensively used technique to process private information over ciphertexts. However, since homomorphic encryption only supports linear operations, approximation techniques are required to calculate nonlinear functions, which leads to the loss of prediction accuracy and heavy computation overhead. Therefore, in this paper, we propose a secure neural network prediction scheme combining the trusted execution environment and homomorphic encryption with different security assumptions. Specifically, we first define the security model of TEE-assisted neural network prediction. Then, by combining a lightweight homomorphic encryption technique with TEE, we design secure neural network prediction protocols under different security levels, with which neural network prediction can be securely processed with high performance and accuracy. Finally, we evaluate the performance of our scheme on the MNIST, Fashion-MNIST, and KMNIST datasets, and the results demonstrate that our scheme indeed improves the prediction efficiency and accuracy compared to traditional homomorphic encryption-based schemes with polynomial approximation.

Today's power systems become more prone to cyber-attacks due to the high integration of information technologies. In this paper, we demonstrate that the outages of some lines can be masked by injecting false data into a set of measurements. The success of the topology attack can be guaranteed by making that: 1) the injected false data obeys Kirchhoff current law and Kirchhoff voltage law to avoid being detected by the bad data detection program in the state estimation and 2) the residual in the line outage detection is increased such that the line outage cannot be detected by phasor measurement unit data. A bilevel optimization problem is set up to determine the optimal attack vector that can maximize the residual of the outaged line. The IEEE 39-bus and 118-bus systems are used to demonstrate the masking scheme.