Health-monitoring apps powered by advanced machine-learning (ML) models could be more secure and still run efficiently on devices, according to researchers from MIT and the MIT-IBM Watson AI Lab. Though the models require vast amounts of data shuttling between a smartphone and a central memory server, using a machine-learning accelerator can speed up the process and reduce the data-transferring need.
However, these accelerators can be vulnerable to cyber-attacks and expose users’ sensitive data. Addressing this concern, the researchers created an ML accelerator that has defenses against the two most prevalent forms of cyberattacks. Despite slightly slowing down the device, this advanced technology does not impact the accuracy of computations and keeps the user data secure.
The chip used in the accelerator can be slightly costlier and less energy-efficient, but according to Maitreyi Ashok, a lead author from MIT, it’s sometimes a necessary trade-off for enhanced security. By designing a system with security in mind from the start, the researchers were able to balance many tradeoffs effectively during the designing phase.
Despite the advantages, in-memory compute (IMC) chips, a type of machine-learning accelerator, can be highly susceptible to hacks. To prevent cyberattacks, the research team adopted a three-pronged approach. They first split data into random pieces, then encrypted the off-chip model with a lightweight cipher, and finally generated a unique key directly on the chip to decrypt the cipher.
To validate their security measures, the researchers attempted to hack into the system and found that they couldn’t reconstruct any real information or extract pieces of the model or dataset. The cipher also remained secure. However, adding extra security measures raised the cost of fabricating the chip and reduced the energy efficiency of the accelerator. The team plans to explore ways to lessen the energy consumption and size of their chip to make it more affordable and scale-ready in the future. The research was funded by the MIT-IBM Watson AI Lab, the National Science Foundation, and a Mathworks Engineering Fellowship and will be presented at the IEEE Custom Integrated Circuits Conference.
