Local Machine Learning Model Fine-Tuning for Privacy Preservation

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing machine learning solutions for personalized user experiences on endpoint devices often compromise user privacy by transferring data to external servers for processing, and they face challenges in achieving real-time, user-specific adaptation on local devices due to processing efficiency and resource utilization limitations.

Innovation Solution

The technique involves fine-tuning a pretrained machine learning model on private customer data using reinforcement learning, all conducted locally on the user device, ensuring that the customer's data never leaves the device. This approach allows the machine learning model to learn and adapt based solely on user-specific data without transferring sensitive information externally.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If data is transferred to external servers for processing, then machine learning model training can be performed, but user privacy is compromised

Engineering Contradiction:

Improvemodel training capabilityVSAvoiduser privacy risk

Core Design Contradiction:

ReliabilityVSObject-affected harmful factors

Solution Approach 1:

Instead of transferring data to external servers for model training, the patent inverts the approach by bringing the model training capability to the local endpoint device. The system performs federated learning and fine-tuning operations locally on the user's device, allowing the model to learn from user data without the data leaving the device, thus maintaining privacy while enabling reliable model training.

Inventive Principle:

Principle #13The other way round (Inversion)

Solution Approach 2:

The patent introduces an intermediary layer in the form of a local model training environment that acts as a mediator between user data and external model services. This local environment enables model training to occur on-device, serving as an intermediary that eliminates the need to transfer sensitive user data to external servers while still providing the necessary training functionality.

Inventive Principle:

Principle #24Intermediary (Mediator)

2Object-affected harmful factors

If machine learning models are trained locally on endpoint devices, then user privacy is preserved, but processing efficiency and resource utilization are limited

Engineering Contradiction:

Improveuser privacy protectionVSAvoidmodel training efficiency

Core Design Contradiction:

Object-affected harmful factorsVSProductivity

Solution Approach 1:

The patent segments the model training process into multiple components that can be executed locally on endpoint devices. By dividing the training workload into manageable tasks that leverage local hardware resources (CPU, GPU, NPU), the system achieves efficient local training without requiring the full processing power of external servers, thus maintaining privacy while improving training efficiency.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent employs parameter changes by adjusting model architecture parameters, training batch sizes, and learning rates to optimize local training efficiency. The system dynamically adapts training parameters based on available local resources, enabling efficient model training on endpoint devices with limited computational power while preserving user privacy.

Inventive Principle:

Principle #35Parameter changes

3Adaptability or versatility

If real-time user-specific adaptation is achieved on local devices, then personalized user experience is improved, but resource utilization challenges increase

Engineering Contradiction:

Improveuser-specific adaptation capabilityVSAvoidresource management complexity

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent implements periodic action by scheduling model fine-tuning and adaptation operations at specific intervals or triggers rather than continuously. The system performs user-specific adaptation periodically based on accumulated user interactions, reducing the continuous resource burden while still achieving effective personalization. This approach balances real-time adaptation capabilities with resource utilization constraints on endpoint devices.

Inventive Principle:

Principle #19Periodic action

Data Source

PatentUS20250094878A1Apparatus for locally training a pretrained machine learning model, a method for locally training a pretrained machine learning model and a non-transitory computer-readable medium

Publication Date: 2025.03.20 INTEL CORP

AI summary

It is provided a non-transitory computer-readable medium storing instructions that, when executed by one or more processing circuitries of an apparatus, causing the one or more processing circuitries to perform locally on the apparatus a method. The method includes obtaining a pretrained machine learning model by the apparatus. The method further includes generating training data based on user-related information. The user-related information relating to a user behavior during interaction of the user with the apparatus. The method further includes training the pretrained machine learning model based on the generated training data by the apparatus to obtain a personalized machine learning model. The method further includes executing the personalized machine learning model by the apparatus.