Enabling privacy-preserving AI training on everyday devices

Enhancing Federated Learning for Edge Devices

MIT researchers have unveiled a breakthrough approach to make privacy-preserving AI training more efficient on devices with limited resources, such as smartwatches and sensors. This innovation enables the use of accurate AI models for critical fields including healthcare and finance—where privacy and security are paramount—even in settings with older or less powerful hardware.

Scaling AI Across a Network of Devices

The team improved upon federated learning, a method where a central server sends an AI model to connected devices. Each device trains the model on its own data and only sends updates—not the original data—back to the server, keeping user information private. However, traditional federated learning faces challenges on devices with limited memory, processing power, and unreliable network connections, slowing down training and reducing overall performance.

Introducing the FTTE Framework

To address these constraints, the researchers developed FTTE (Federated Tiny Training Engine), introducing three key strategies:

• Selective Parameter Updates: Only a targeted subset of model parameters, rather than the entire model, is distributed to each device, minimizing memory use.
• Asynchronous Communication: The server processes updates as they arrive without waiting for all devices, which reduces idle time and accelerates training.
• Weighted Updates: Updates are weighted based on their arrival time—newer updates contribute more, ensuring the model remains current and efficient.

Impressive Results and Real-world Impact

Simulations involving hundreds of diverse devices and a range of models showed that FTTE speeds up AI training by approximately 81% compared to standard approaches. It also cuts device memory requirements by 80% and reduces communication loads by 69%, with only a minor decrease in accuracy. These gains make it possible to deploy advanced AI on a wider variety of personal devices, expanding access for users in under-resourced regions or with older technology.

Looking ahead, the researchers plan to refine their approach for personalized AI models and expand testing on real-world hardware to further enhance its applicability.

For more details, read the original article at MIT News.