Assisted Learning Framework for Privacy-Preserving Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional machine learning architectures face challenges in balancing data privacy and accuracy, often requiring the disclosure of user data, which poses security risks and compliance issues with regulations like GDPR, and there is a need to protect both data and proprietary models from adversaries.
Innovation Solution
The proposed solution involves an assisted learning framework where agents build local models and share intelligence without disclosing sensitive data, using a communication protocol that protects both learner units and benefits their learning performance, enabling data and algorithm privacy through iterative exchanges of statistical information, such as confidence scores, within a network of remote computing devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If user data is disclosed to train machine learning models, then model accuracy is improved, but data security and privacy are compromised
Solution Approach 1:
The patent extracts only the essential statistical information (confidence scores, gradient statistics) from the training data and model, sharing these extracted features between agents while keeping the raw user data private and local to each agent. This allows model improvement without exposing sensitive data.
Solution Approach 2:
The patent introduces an intermediary communication protocol that mediates between data privacy requirements and model accuracy needs. The protocol enables agents to share statistical information through encrypted or privacy-preserving channels, acting as a mediator that allows collaboration while maintaining security boundaries.
2Measurement precision
If user data is disclosed for model training, then learning quality is improved, but compliance with regulations like GDPR is violated
Solution Approach 1:
The patent extracts anonymized statistical features (confidence scores, gradient magnitudes) from the training process, which do not contain personally identifiable information. This extraction approach maintains learning quality while ensuring GDPR compliance by removing direct data exposure.
Solution Approach 2:
The patent creates copies of statistical information (confidence scores, gradient statistics) that can be shared between agents without copying or exposing the original user data. These statistical copies enable collaborative learning while maintaining data privacy and regulatory compliance.
3Adaptability or versatility
If proprietary models are shared for assisted learning, then collective intelligence is improved, but model privacy and security are compromised
Solution Approach 1:
The patent extracts only the necessary statistical components (confidence scores, gradient information) from proprietary models for sharing, while keeping the core model architecture and proprietary algorithms private. This selective extraction enables collective intelligence improvement without exposing sensitive model properties.
Solution Approach 2:
The patent introduces a privacy-preserving communication protocol as an intermediary that enables model assistance sharing without direct model exposure. The protocol mediates the interaction between agents, allowing statistical information exchange while protecting proprietary model security.
Data Source
AI summary
Techniques are disclosed for assisted learning with enhanced privacy. A method comprises: sending first statistical information from a first agent to a second agent in an architecture having at least two agents, wherein a first set of sample weights correspond to training the first machine learning model, wherein the first statistical information comprises the second set of sample weights determined from a first model weight; receiving, from the second agent, second statistical information comprising the second model weight and updated first set of sample weights or, from a third agent of the architecture, third statistical information comprising a third model weight and a next iteration of the first set of sample weights; and updating the first machine learning model using the second statistical information or the third statistical information.