How Devices Securely Share Machine Learning Model Updates

This patent describes a method for multiple devices to train a machine learning model using their own data and then securely share parts of the trained model with each other before sending it to a central server, improving data privacy.

What it covers

This patent details a process for training machine learning models across many devices, like smartphones or computers, without sending all the raw data to a central server. A 'first terminal' (one of the devices) gets a model from a server, trains it using its own local data to create 'trained model parameters,' and then figures out which other devices ('first collaborative terminal') it should work with based on a 'collaboration relationship' provided by the server. It then splits its trained model parameters, encrypts parts of them with 'first encryption noise,' and sends these encrypted parts to its collaborative terminal(s). The collaborative terminal(s) also encrypt their received parts with 'second encryption noise.' Finally, these encrypted pieces are sent to the main server, which combines them to build a better, more accurate machine learning model. This approach aims to make data exchange more secure.

What it doesn't cover

• —Does not cover methods where devices send all their local data to the server.
• —Does not cover training methods that do not involve a server coordinating the process.
• —Does not cover scenarios where devices do not identify or collaborate with other specific devices for parameter sharing.
• —Does not cover methods where the model parameters are not encrypted with noise before transmission.
• —Does not cover systems where the 'collaboration relationship' is not provided by the server.

The clever bit

The innovation lies in the multi-stage encryption and collaborative sharing of model parameters between devices before they reach the central server. By splitting parameters and encrypting them with different 'noise' at both the originating and collaborative terminals, it creates a more robust privacy shield than simply encrypting once.

Why it matters

This patent is part of the growing field of federated learning, which is crucial for training AI models on sensitive data like personal health information or user behavior without compromising privacy. It offers a technical solution to enhance the security of data exchange in such distributed learning systems.

Real-world examples

• 1.Secure training of AI models on mobile devices
• 2.Collaborative machine learning for smart home devices
• 3.Privacy-preserving AI for healthcare applications
• 4.Federated learning systems for edge computing