Researchers from the Massachusetts Institute of Technology (MIT) and the MIT-IBM Watson AI Lab have developed a machine-learning accelerator designed to be resistant to cyber-attacks, offering a secure platform for health-monitoring applications. The chip secures users’ data whilst running large artificial intelligence (AI) models efficiently, protecting sensitive health and financial information.
The technology is capable of resisting the two most common types of cyber-attacks, using three pronged strategies. The first projects data in the chip that is divided into random parts, and are encrypted utilizing low-weight ciphers that only require simple computations. The encryption key is generated directly on the device by the physically unclonable function using natural chip variations produced during manufacturing. The key is stored on the device, eliminating the need to be transferred with the data model. To ensure the security measures are foolproof, the researchers attempted to hack into the device using side-channel and bus-probing attacks in a hostile environment. These test results demonstrated the new security system’s effectiveness. The researchers were unable to reconstruct any real data or model segments, and the cipher remained unbroken even after millions of attempts.
Despite the advances, the security measures have taken a toll on both the size and energy efficiency of the hardware. The additional encryption capabilities enlarged the chip’s fabrication size, impacting the cost and energy consumption. The developers are looking at future potentials to shrink the size and reduce the energy expended by the chip without compromising the security. Lead author and EECS graduate student at MIT, Maitreyi Ashok, acknowledges that security generally comes at a cost, yet foresees that this may be a balancing act in convincing future developers of the importance of a robust security network.
MIT’s Chief Innovation and Strategy Officer, dean of the School of Engineering, and the Vannevar Bush Professor of EECS, Anantha Chandrakasan, expressed the need for continued research emphasis on secure edge devices, particularly focusing on machine-learning workloads. The research, funded, in part, by the MIT-IBM Watson AI Lab, National Science Foundation, and the Mathworks Engineering Fellowship, revealed at the IEEE Custom Integrated Circuits Conference, is set to play a critical role in the security of future mobile devices.