Secure Neural Network Inference for Edge Intelligence: Implications of Bandwidth and Energy Constraints

Recently, there has been a growing interest in machine learning as a service (MLaaS). In MLaaS, an operator provides pretrained neural networks, with which inferences on clients’ inputs can be performed. MLaaS is attractive for providing edge intelligence in an Internet of Things (IoT) setup for multiple reasons, for example, because it relieves clients with limited capacity from the computationally heavy training process. However, MLaaS may lead to privacy threats for both the client and the provider. In particular, the input of the client may be sensitive information that the provider is not allowed to learn. The provider, on the other hand, may not want to reveal the parameters of the neural network to the client, because these parameters are the provider’s intellectual property. Besides, the output of the neural network might also reveal sensitive properties about the input. Lastly, traditional security solutions might fail in an IoT setup. In recent years, several cryptographic protocols have been devised for secure neural network inference (SNNI). Secure neural network inference entails the problem of computing the output of a neural network on the client’s input without revealing the input to the provider or the parameters of the neural network to the client. So far, SNNI approaches have been optimized for efficiency and accuracy, mainly in cloud settings. The goal of this chapter is to investigate the applicability of SNNI approaches in an edge computing setup. In particular, with power-constrained edge and IoT devices in mind, we investigate the power consumption and energy consumption characteristics of SNNI approaches. Taking into account the typical bandwidth of access networks relevant to edge and IoT, we also investigate the effect of bandwidth limitations on the duration and energy consumption of the SNNI process. Our results indicate that the power consumption of SNNI depends significantly on both the used SNNI protocol and the available bandwidth.