How Distributed AI Models Learn Privately on Edge Devices

This patent describes how a central computer can train a basic artificial intelligence model, send it to a remote device, and let that device customize the model using its own private data, then use it to filter information before sending it back.

What it covers

This patent details a system where a central computer, called a "first computing system," first trains a neural network using general data collected from many remote devices (claim 1). Once this initial training is done, the central system sends this partially trained neural network to a specific remote device, a "second computing system." This remote device then takes the neural network and further trains it using its own "private data" that isn't shared with the central system (claim 1). For example, a security camera (the second computing system) could receive a general object detection model, then train it specifically on its own footage of common local events, like a delivery person, without sending that private footage back to the central server. The customized neural network then processes "live event data" directly on the remote device, filtering out less important events before any data is sent back to the central system (claim 1, claim 4).

What it doesn't cover

• —Does not cover a system where all neural network training, including customization, happens only on a single central server.
• —Does not cover scenarios where the remote device sends all its raw private data back to the central system for training.
• —Does not cover a system where the customized neural network is not applied to filter live event data at the remote device.
• —Does not cover a system where the central system does not initially train a neural network before sending it to the remote system.
• —Does not cover a system where the remote device's private data is not used to further train and customize the neural network.

The clever bit

The novelty lies in the specific two-stage training process: a general model is trained centrally, then sent to an edge device for private, localized refinement. This allows for personalized AI performance while keeping sensitive data on the device where it originates, rather than sending it all to a central server.

Why it matters

This technology is important for applications where data privacy is critical, such as surveillance, healthcare, or industrial monitoring. By allowing AI models to be customized on local devices using private data, it reduces the need to transmit sensitive information to a central server. This approach can also make AI systems more efficient by filtering irrelevant data at the source, saving bandwidth and processing power for the central system.

Real-world examples

• 1.Smart security cameras that learn to recognize specific local events without uploading all footage.
• 2.Industrial sensors that adapt anomaly detection models to unique machine behaviors on-site.
• 3.Personalized health monitoring devices that refine AI models using individual user data.
• 4.Federated learning systems where models are trained collaboratively without sharing raw data.