Enabling privacy-preserving AI training on everyday devices

Source Domain: news.mit.edu

MIT researchers have developed a new method that accelerates a privacy-preserving AI training method by about 81 percent, enabling resource-constrained edge devices to deploy more accurate AI models while securing user data.
The advancement involves optimizing federated learning, a technique where a shared AI model is trained collaboratively across a network of edge devices using their local data while keeping it secure.
The new approach, FTTE (Federated Tiny Training Engine), addresses the challenges faced by wireless devices with varying memory and connectivity, mitigating memory constraints and communication bottlenecks.
FTTE’s main innovations include sending smaller subsets of model parameters to devices, an asynchronous updating approach by the server to optimize resource usage, and a weighting system for model updates based on their recency.
In tests, FTTE achieved training speeds 81 percent faster than standard federated learning, reduced on-device memory usage by 80 percent, and decreased communication payload by 69 percent, while maintaining near comparable accuracy.
This improvement is critical for deploying AI on everyday edge devices like sensors and smartwatches, especially in sectors requiring high security and privacy standards, such as healthcare and finance.

Enabling privacy-preserving AI training on everyday devices | MIT News