Automatic Unlearning Analysis for Data Classification Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data classification models trained using machine learning lack effective methods to verify the unlearning of sensitive classes, and there is a need to ensure that sensitive class instances are not used in third-party models without knowledge of the initial training data.
Innovation Solution
A method and device for automatically analyzing unlearning in data classification models by calculating homogeneity values and re-learning characteristics to determine the probability that candidate classes are forgotten classes, using witness models and a final score to validate unlearning effectiveness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If unlearning is performed to remove sensitive classes from a classification model, then security is improved, but verification of unlearning effectiveness becomes difficult without access to initial training data
Solution Approach 1:
The patent introduces witness models as intermediary systems that do not require access to initial training data. These witness models are trained on available data and serve as reference points to compare against the target model's behavior on candidate forgotten classes, enabling verification through indirect comparison rather than direct access to original training materials.
Solution Approach 2:
The patent creates copies of classification behavior through witness models that replicate the structure and training process of the target model. By generating multiple witness models with known training histories, the system can compare their responses to candidate classes against the target model's responses, copying the verification function without requiring access to the original sensitive training data.
2Measurement precision
If multiple witness models are trained to verify unlearning, then verification accuracy is improved, but computational complexity and training time increase
Solution Approach 1:
The patent trains witness models on partial subsets of available training data rather than requiring complete datasets. By using representative samples and focusing training on specific aspects of class characteristics, the system achieves sufficient verification accuracy without the excessive computational burden of training on all possible data, applying partial action to reach adequate verification levels.
3Measurement precision
If homogeneity analysis and re-learning characteristics are both used to determine forgotten classes, then detection accuracy is improved, but processing time and computational resources increase
Solution Approach 1:
The patent segments the verification process into two independent analysis streams: homogeneity analysis of latent space representations and re-learning characteristic analysis. These segmented approaches can be applied separately or in combination, allowing the system to process different aspects of verification independently and combine results, improving detection accuracy through multiple perspectives while managing computational load through modular processing.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present invention relates to a method and a device for automatic analysis of unlearning of at least one class by a target model (MC) for data classification, among a set of candidate classes (20). The device implements a plurality of witness models (MT1, ..., MTN) learned by machine learning, and modules for: first determination (32), as a function of homogeneity values calculated in the latent space (12, 121...12N) for the target model and for each witness model, providing as output a first binary prediction and a first associated prediction score, and/or second determination (34), as a function of a re-learning characteristic of refinement of additional classes among said candidate classes by the target model and the witness models, providing as output a second binary prediction and associated prediction score, and calculation (36) of a final score as a function of the first and second prediction scores.