The outcome is for more efficient security for cloud-based machine learning. The approach comes from the Massachusetts Institute of Technology and it is focused with securing data used in online neural networks. A secondary brief was to boost security while also avoiding significantly slowing down machine runtimes. A problem with many cybersecurity solutions is that they tend to slowdown the very device they aim to protect.
The harnessing of machine learning with the cloud is important since more organizations are outsourcing machine learning. To meet this demand, most of the leading technology companies have developed cloud platforms that can conduct computation-heavy tasks, such as running data through a convolutional neural network (a class of deep, feed-forward artificial neural network) for image classification. This type of approach is popular with many medical applications.
To test out the new approach, the researchers used their new system which is called GAZELLE. This was used to assess and protect several two-party image-classification tasks. With this, a user sends an encrypted image data to an online server evaluating a convolutional neural network running on GAZELLE. Following this, both parties share encrypted information back and forth.
By using the new method, GAZELLE ensures that the server does not ‘learn’ any uploaded data and that the user never learns anything about the network parameters. The outcome of the research is a viable approach for using cloud-based neural networks for medical-image analysis and other applications that use sensitive data, by affording them enhanced security without impairing device operations.
The research was presented to the August 2018 USENIX Security Conference. The next step is to further develop the GAZELLE platform and offer it for commercial use. This could lead to more uses of Internet-based neural networks for handling vital information.
