New federated learning approach highlights shift towards decentralised and privacy-preserving AI

Researchers at MIT have developed a new method that significantly improves privacy-preserving AI training on everyday devices such as smartphones, sensors, and smartwatches.

The approach strengthens federated learning systems, where data remains on devices while models are trained collaboratively, supporting sensitive applications such as healthcare and finance.

The new framework, called FTTE (Federated Tiny Training Engine), addresses long-standing issues in federated learning networks with uneven device capabilities. Traditional systems struggle with delays from limited memory, weak connectivity and slow update cycles, reducing network efficiency and performance.

FTTE improves the process by sending smaller model segments to devices, introducing asynchronous updates and weighting contributions based on freshness. These changes reduce memory load and communication demands while maintaining stable training across heterogeneous devices.

Testing across simulated and real device networks showed training speeds improved by around 81 percent, with major reductions in memory and data transfer requirements.

Researchers also highlighted the potential to expand AI access in regions with lower-end hardware, while future work will focus on further personalising models for individual devices.

Why does it matter? 

Decentralised AI training marks a shift away from dependence on centralised data centres towards distributed intelligence embedded in everyday devices.

That changes the architecture of AI itself, allowing sensitive data to remain local and reducing privacy risks. At the same time, computation is spread across billions of low-power devices rather than concentrated in a few powerful systems.

The researchers note that such approaches may enable AI training on devices with limited memory and connectivity.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!