User Information Purge in Learning Devices via Selective Forgetting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for protecting user privacy in systems processing user information, such as those in the 'Internet of Things' and healthcare, often result in loss of precision due to suppression or deletion of data, and existing anonymization techniques fail to effectively manage user information while maintaining privacy.
Innovation Solution
A method and device that purge user information from machine learners and user models by providing only data that adheres to predefined rules, allowing for selective deletion and retraining of models to suppress the influence of user information, using techniques like concept drift correction and ensemble methods to adapt machine learners and user models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If user information is suppressed or deleted before processing by a machine learner, then user privacy is protected, but the precision and accuracy of the machine learner deteriorates
Solution Approach 1:
The patent segments the machine learning process into multiple stages: initial training with user information, operation phase, and forgetting phase. The user information is processed in a controlled manner through these segments, allowing privacy protection at specific stages while maintaining learning precision during others. The forgetting process itself is segmented into identifying, training, and evaluating sub-stages.
Solution Approach 2:
The patent applies preliminary action by pre-training the machine learner with user information before actual operation, then preparing a forgetting mechanism in advance. The system pre-identifies which user information should be forgotten and pre-trains alternative models, so that when forgetting is needed, the transition is smooth and precision is maintained.
2Object-affected harmful factors
If user information is anonymized with random data, then user privacy is protected, but the ability to associate findings with users is lost
Solution Approach 1:
The patent implements dynamic control over user information retention and forgetting. Instead of static anonymization, the system dynamically adjusts what user information is retained or forgotten based on user preferences, regulatory requirements, and operational context. This allows flexible association capability while maintaining privacy protection when needed.
Solution Approach 2:
The system changes parameters of user information processing by adjusting the forgetting factor, training data composition, and model architecture parameters. These parameter changes allow the system to transition between states of high user association capability and high privacy protection without permanent loss of information.
3Object-affected harmful factors
If distributed machine learners are used to generate encrypted results, then user privacy is protected, but system complexity increases
Solution Approach 1:
The patent introduces a central management unit as an intermediary that coordinates distributed machine learners. This intermediary handles the complex tasks of distributing forgetting instructions, collecting results, and managing the overall forgetting process, thereby reducing the complexity burden on individual distributed learners while maintaining privacy protection.
4Object-affected harmful factors
If machine learners are completely retrained to remove user information influence, then user information is purged from models, but training time and computational resources increase
Solution Approach 1:
The patent applies partial action by implementing selective forgetting rather than complete retraining. The system identifies and forgets only the specific user information that needs to be removed, while retaining other useful learned patterns. This partial forgetting approach significantly reduces training time and computational resources compared to complete retraining, while still achieving the goal of purging unwanted user information.
Solution Approach 2:
The system discards specific user information influences from the machine learner through targeted forgetting processes, while recovering and retaining other valuable learned patterns. This selective discarding and recovering mechanism allows efficient removal of unwanted information without losing overall model competence.
Data Source
AI summary
A method in which user information is transmitted from at least one data source to a processing unit of a learning device. The user information is used, by the processing unit via a machine learner, to generate at least one user model. The at least one user model is adapted via an adaptation of parameters used by the at least one machine learner to generating the at least one user model. The parameters, used by the at least one machine learner for generating the at least one user model, are adapted as a function of at least one predefined rule. The user model generated on the basis of the adapted parameters is used to personalize at least one terminal.