MIT Secures Cloud-Based AI With Novel 20 to 30 Times Faster Method
The use of public clouds is on the rise with advisory firm Gartner forecasting that a whopping $186.4 billion will be spent on the services globally in 2018. Tech giants such as Amazon, Google and Microsoft have even launched cloud-based artificial intelligence (AI) platforms capable of conducting computation-heavy tasks through the use of convolutional neural networks.
However, with great power comes great responsibility. As the cloud becomes ever more useful in ever more applications the potential for security breaches also increases.
Security breaches looming
A study conducted in 2016 by McAfee-acquired security firm Skyhigh Networks on 30 million of its software users revealed that an average enterprise experiences an alarming 23.2 cloud-related threats per month. Researchers in the past few years have explored a variety of secure-computation techniques for cloud-related applications.
However, when it comes to cloud-based AI platforms, attempting to encrypt the stored data has so far rendered the systems so painfully slow they become unusable. Now, a novel encryption method has been developed by MIT researchers capable of securing online neural networks without dramatically slowing them down.
The new security system consists of the combination of two conventional encryption techniques, homomorphic encryption and garbled circuits, devised in such a way that it circumvents their individual inherent inefficiencies. “We’re only using the techniques for where they’re most efficient,” explained first author Chiraag Juvekar, a PhD student in the Department of Electrical Engineering and Computer Science.
The researchers called their ingenious new system GAZELLE, in reference to its unique ability to protect data while keeping neural networks running quickly. They then proceeded to test GAZELLE on two-party image-classification tasks.
The process functioned efficiently protecting both the uploaded data and the network's parameters just as a traditional system would. However, GAZELLE ran 20 to 30 times faster than even state-of-the-art models.
The innovative system has countless promising and exciting applications. Perhaps the most notable is the role it could one day play in the medical field.
Sharing medical data securely
“The next step is to take real medical data and show that, even when we scale it for applications real users care about, it still provides acceptable performance," said Juvekar. One such example would see convolutional neural networks trained by hospitals to recognize medical condition characteristics from magnetic resonance images.
Using GAZELLE, hospitals could then securely and efficiently share their results in the cloud with other healthcare institutions, furthering medical advancements while protecting private patient data. If such a feat seems too good to be true, you may be reassured to know it is not.
Cloud use has become so pervasive that, according to a 2016 annual cloud computing survey by venture firm North Bridge, 50% of organizations followed either a cloud-first or cloud-only policy while a staggering 90% were using the cloud in some way. With GAZELLE now enabling secure cloud-based machine learning systems, there is no telling how ubiquitous cloud platforms will become.
A team of researchers in the U.S. and China have developed a new paradigm for enabling communication between humans and AI systems.