In Short:
Health-monitoring apps on smartphones help manage chronic diseases and fitness goals but can be slow and vulnerable to attacks. Researchers from MIT and MIT-IBM Watson AI Lab created a secure machine-learning accelerator chip resistant to common attacks, ensuring data privacy. While chip adds cost and reduces energy-efficiency, ensuring security from the start is crucial. Testing showed security measures successfully blocked attacks. Future research aims to improve efficiency and cost-effectiveness.
MIT Researchers Develop Secure Machine-Learning Accelerator for Health-Monitoring Apps
Health-monitoring apps have become essential tools for managing chronic diseases and fitness goals, using smartphones for convenience. However, the performance of these apps is often hindered by slow and energy-inefficient processes due to large machine-learning models that need to be transferred between smartphones and central memory servers.
Ensuring Security in Machine-Learning Accelerators
Engineers typically use hardware to accelerate computation and reduce data movement, but these accelerators are vulnerable to attacks that compromise sensitive information. To address this, researchers from MIT and the MIT-IBM Watson AI Lab developed a secure machine-learning accelerator resistant to common attacks.
The chip created by the researchers provides privacy for user data while efficiently running large AI models on devices, such as health records and financial information.
Preventing Side-Channel Attacks
The researchers focused on digital in-memory compute (IMC) chips, which perform computations within a device’s memory. By splitting data into random pieces and using encryption techniques, the chip effectively blocks side-channel and bus-probing attacks, maintaining the security of user information.
Future Implications and Security Testing
The secure chip, while slightly more expensive and less energy-efficient, provides significant security benefits. In security tests, the researchers were unable to extract real information even after multiple attempts, highlighting the chip’s robust defenses against attacks.
While the chip’s implementation may have trade-offs in cost and efficiency, the team plans to explore ways to enhance energy consumption and chip size for broader implementation.
The project is supported by funding from the MIT-IBM Watson AI Lab, the National Science Foundation, and a Mathworks Engineering Fellowship.
