Enabling privacy-preserving AI training on everyday devices

Source: MIT News - AI

Accelerating Privacy‑Preserving AI Training by 81 %

A new method developed by MIT researchers can accelerate a privacy‑preserving artificial‑intelligence training method by about 81 percent. This advance could enable a wider array of resource‑constrained edge devices—such as sensors and smartwatches—to deploy more accurate AI models while keeping user data secure.

Federated Learning Overview

Federated learning involves a network of connected devices that collaboratively train a shared AI model:

1. The model is broadcast from a central server to wireless devices.
2. Each device trains the model using its local data.
3. Devices send model updates back to the server.

Because the raw data never leave the devices, user privacy is preserved.

The Problem

Many devices lack the capacity, computational capability, and connectivity needed to store, train, and transfer the full model quickly. These constraints cause delays that degrade training performance.

MIT’s Solution

The MIT team created a technique to overcome memory constraints and communication bottlenecks, designed for a heterogeneous network of wireless devices with varied limitations. This approach could make AI models more feasible for high‑stakes, privacy‑sensitive applications such as health care and finance.

“This work is about bringing AI to small devices where it is not currently possible to run these kinds of powerful models. We carry these devices around with us in our daily lives. We need AI to be able to run on these devices, not just on giant servers and GPUs, and this work is an important step toward enabling that,”
— Irene Tenison, EECS graduate student and lead author of a paper on this technique.

Co‑authors

• Anna Murphy ’25 – Machine‑learning engineer, Lincoln Laboratory
• Charles Beauville – Visiting student, EPFL (Switzerland) & machine‑learning engineer, Flower Labs
• Lalana Kagal – Principal research scientist, CSAIL, MIT (senior author)

The research will be presented at the IEEE International Joint Conference on Neural Networks.

---

Reducing Lag Time

Many federated‑learning approaches assume every device can:

• Store the full AI model
• Maintain stable connectivity to transmit updates quickly

These assumptions break down for heterogeneous edge devices (smartwatches, sensors, mobile phones) that have limited memory, compute power, and intermittent connectivity.

The central server traditionally waits for updates from all devices, averages them, and repeats the process—leading to lag that can slow or even halt training.

“This lag time can slow down the training procedure or even cause it to fail,” — Tenison

FTTE: Federated Tiny Training Engine

To address these issues, the MIT researchers introduced FTTE, a framework that reduces both memory and communication overhead per device. FTTE comprises three main innovations:

1. Partial Model Broadcast

   • Instead of sending the entire model, FTTE transmits a smaller subset of parameters tailored to the most memory‑constrained device.
   • A specialized search procedure selects parameters that maximize accuracy while respecting a predefined memory budget.

2. Semi‑Asynchronous Server Updates

   • The server does not wait for every device. It accumulates incoming updates until a fixed capacity is reached, then proceeds with the training round.

3. Weighted Updates by Recency

   • Updates are weighted based on arrival time; older updates contribute less, preventing stale data from dragging down accuracy.

“We use this semi‑asynchronous approach because we want to involve the least powerful devices in the training process so they can contribute their data to the model, but we don’t want the more powerful devices in the network to stay idle for a long time and waste resources,” — Tenison

---

Achieving Acceleration

• Simulation Results: Tested with hundreds of heterogeneous devices across various models and datasets.
  • Training completed 81 % faster than standard federated learning.
  • On‑device memory overhead reduced by 80 %.
  • Communication payload cut by 69 %.
  • Accuracy remained near‑par with existing techniques.

“Because we want the model to train as fast as possible to save the battery life of these resource‑constrained devices, we do have a trade‑off in accuracy. But a small drop in accuracy could be acceptable in some applications, especially since our method performs so much faster,” — Tenison

• Scalability: FTTE showed larger performance gains as the number of devices increased.
• Real‑World Testbed: Deployed on a small network of devices with varying computational capabilities.

“Not everyone has the latest Apple iPhone. In many developing countries, for instance, users might have less powerful mobile phones. With our technique, we can bring the benefits of federated learning to these settings,” — Tenison

Future Directions

The team plans to explore how FTTE can personalize AI performance on each device, shifting focus from average performance to individualized optimization.

of the model. They also want to conduct larger experiments on real hardware.

This work was funded, in part, by a Takeda PhD Fellowship.