Differential AI Model Un-learning via Feature Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for un-learning user data from AI models are cumbersome, difficult to scale, and costly, especially in enterprise-wide implementations, as they require re-training and re-deploying models, which can cause poor customer experiences due to frequent updates and are unscalable with the current proliferation of AI models deployed across millions of systems.
Innovation Solution
The system employs a differential AI process to generate features and models without individual-specific data, allowing for the comparison of native and differential features and models to determine influence levels, enabling efficient un-learning by replacing native features and models with differential ones, thus allowing for quick and efficient removal of individual data without affecting ongoing learning processes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional re-training and re-deployment methods are used to remove user data from AI models, then user privacy is protected, but system productivity and scalability deteriorate due to frequent updates and high costs
Solution Approach 1:
The patent segments the AI model into two distinct components: native features/models trained on all data including individual user data, and differential features/models trained only on aggregated data excluding individual user data. This segmentation enables selective replacement of only the differential component when un-learning is needed, rather than re-training the entire model, thus maintaining scalability and productivity while protecting user privacy.
Solution Approach 2:
The patent performs preliminary action by pre-computing and storing differential features and models before they are needed for un-learning operations. When a user requests data removal, the system can immediately swap in the pre-computed differential model without initiating time-consuming re-training processes, thereby maintaining high productivity and scalability.
2Reliability
If native AI models are re-trained to remove individual data, then privacy is maintained, but time consumption increases due to the re-training process
Solution Approach 1:
The system performs preliminary action by pre-computing differential features and models in advance, storing them for immediate use. When un-learning is required, the pre-computed differential model can be swapped in without time-consuming re-training, significantly reducing the time loss while maintaining privacy through the differential modeling approach.
Solution Approach 2:
The patent extracts the individual user data influence from the native model by creating a separate differential model that explicitly excludes individual data. This extraction allows the system to maintain privacy by separating individual contributions from the collective model, and enables quick replacement of only the extracted differential component without re-training the entire model.
3Adaptability or versatility
If frequent model updates are performed to remove user data, then user rights are protected, but customer experience deteriorates due to service disruptions
Solution Approach 1:
By segmenting the model into native and differential components, the system can update only the differential portion when user rights compliance is needed, rather than performing full model re-training and re-deployment. This segmented update approach minimizes service disruptions and maintains customer experience while still protecting user rights.
Solution Approach 2:
The system performs preliminary action by pre-computing differential models that can be immediately deployed when needed. This eliminates the need for time-consuming re-training processes that would disrupt service, allowing the system to comply with user rights requirements while maintaining seamless customer experience through instant model swapping.
4Reliability
If individual data is completely removed from training sets, then privacy is enhanced, but model accuracy may deteriorate due to loss of valuable training information
Solution Approach 1:
The patent segments the modeling process into native features (trained on all data including individual data) and differential features (trained on aggregated data excluding individual data). This segmentation allows the native model to retain full training information for accuracy, while the differential model provides the privacy-protecting mechanism. Both components work together to achieve both privacy enhancement and model accuracy.
Solution Approach 2:
The patent applies parameter changes by modifying the training data composition for the differential model rather than removing data from the native model. The differential model uses aggregated parameters and features that exclude individual data, while the native model maintains its full training set. This parameter change approach preserves valuable training information in the native model while achieving privacy enhancement through the differential component.
Data Source
AI summary
According to one illustrative, non-limiting embodiment, an IHS may include computer-executable instructions for receiving individual-specific raw data instances associated with an individual, and other raw data instances that are not associated with the individual to perform a native artificial intelligence (AI) process to generate one or more native features according to the received other raw data instances and the individual-specific raw data instances, and a native AI model from the one or more native features. The instructions also perform a differential AI process to generate one or more differential features according to the received other raw data instances, and a differential model from the one or more differential features. The native features may be compared against the differential features to determine one or more feature influence levels, while the native model may be compared against the differential model to determine a model influence level of the individual-specific raw data instance on the native model.