Health-monitoring apps are widely used today to help manage chronic diseases and fitness goals, but they often face challenges related to speed and energy-efficiency due to the large machine-learning models powering them being moved between smartphones and central memory servers. To address this issue, researchers from MIT and the MIT-IBM Watson AI Lab have developed a machine-learning accelerator that enhances security while maintaining efficient performance. This chip can protect sensitive data such as health records and financial information while allowing for the smooth operation of AI models on devices.
The new machine-learning accelerator is resistant to common types of attacks, making it particularly useful for demanding AI applications like augmented reality, virtual reality, and autonomous driving. Though implementing the chip may increase the cost and reduce energy-efficiency of a device, it is a small price to pay for added security, according to lead author Maitreyi Ashok, an electrical engineering and computer science graduate student at MIT. It is crucial to design with security in mind from the beginning to avoid costly overheads later on.
The researchers focused on a type of machine-learning accelerator called digital in-memory compute, which performs computations inside a device’s memory, reducing the amount of data that needs to be shuttled back and forth. However, these chips are vulnerable to side-channel attacks where hackers can monitor power consumption or steal data using bus-probing techniques. To combat these threats, the researchers took a three-pronged approach to enhance security, including splitting data into random pieces, encrypting the model stored in off-chip memory, and generating unique keys on the chip using physically unclonable functions.
Testing their chip as hackers, the researchers found that even after millions of attempts, it was impossible to extract real information or break the encryption. This demonstrates the effectiveness of their security measures in preventing unauthorized access to sensitive data. While the added security did impact energy efficiency and chip size, the researchers plan to explore ways to optimize these aspects in future iterations to make the chip more scalable and cost-effective.
The research is supported by the MIT-IBM Watson AI Lab, the National Science Foundation, and a Mathworks Engineering Fellowship, indicating the significance of this work in advancing secure AI technologies for mobile devices. As security continues to be a critical concern in edge devices, the development of secure systems focusing on machine-learning workloads is essential. The researchers believe that their design choices, such as encrypted data access and protection against side-channel attacks, will be crucial in improving security in mobile devices and facilitating secure operation in the future.
