Deep neural network obfuscator for machine learning as a service in presence of cache side-channel attacks

Abstract

As Deep Neural Networks (DNNs), one of the machine learning algorithms, has solved many complex problems with good performance, the demand of DNNs has increased. DNNs are provided to users in the form of Machine Learning as a Service (MLaaS) because of its large computational complexity, and service providers are obliged to design and analyze the neural networks with good performance and provide them to users. Therefore, a good performance DNNs has high commercial value. For this reason, researches have been published to obtain architectural information of DNNs using cache side-channel attacks in the cloud environment. In this dissertation, we introduce several mitigation techniques to prevent such attacks and analyze their effects. Also, we propose an obfuscator that conceals the dimension of each layer which is one of the architectural information of DNNs. This obfuscator hides the real dimension value from the attacker by making all the dimensions of each layer of the neural network equal. Finally, we optimize the performance of the obfuscator in a way that does not significantly degrade obfuscation and evaluate inference time, memory usage, and side-channel vulnerability metrics of obfuscated DNNs.