Personal Data Discovery in Machine Learning Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current technologies face challenges in identifying and managing personal data within machine learning models, particularly in adhering to regulations like GDPR and CCPA, as malicious third-parties can reveal sensitive information without access to training data, and there is a need to detect personal data leakage in ML models.
Innovation Solution
A method and system for personal data discovery that uses data discovery tools, supervised or unsupervised ML techniques, and adversarial inference attacks to locate and quantify personal data in data stores and ML models, employing membership leakage metrics to determine if the model contains or leaks personal data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If adversarial inference attacks and membership leakage metrics are applied to detect personal data in ML models, then personal data detection capability is improved, but computational complexity and processing time increase
Solution Approach 1:
The system performs preliminary classification of data stores to identify which ones contain personal data before applying complex adversarial inference attacks. This preliminary sorting action reduces the number of targets requiring intensive computational analysis, thereby maintaining high detection capability while reducing overall computational complexity.
Solution Approach 2:
The detection process is segmented into multiple stages: initial data store classification, followed by targeted adversarial attacks only on suspicious models, and finally membership leakage metric calculation only for those cases. This segmentation allows the system to achieve high detection precision while distributing computational load across manageable segments rather than applying all methods uniformly to all data.
2Reliability
If comprehensive personal data discovery is performed across all data stores and ML models, then data security and privacy are improved, but processing time and computational resources increase
Solution Approach 1:
The system performs preliminary classification of data stores using supervised or unsupervised ML techniques to identify those containing personal data before conducting comprehensive analysis. This preliminary action filters out non-sensitive data stores, allowing comprehensive personal data discovery to be focused only on relevant targets, thereby maintaining high data security while reducing processing time.
Solution Approach 2:
The system applies adversarial inference attacks and membership leakage metrics selectively to data stores and ML models that are identified as containing personal data, rather than applying these computationally intensive methods universally. This partial action approach ensures thorough security checking where needed while avoiding unnecessary processing time on non-sensitive data.
3Measurement precision
If ML techniques are used to classify and locate personal data in data stores, then personal data identification accuracy is improved, but system complexity increases
Solution Approach 1:
The system employs universal ML classification techniques that can operate in both supervised and unsupervised modes, allowing the same core technology to handle various types of personal data across different data store formats and structures. This multi-functionality approach improves identification accuracy across diverse scenarios while avoiding the need for multiple specialized systems, thereby controlling overall system complexity.
Data Source
AI summary
A method, computer system, and a computer program product for personal data discovery is provided. The present invention may include determining at least one feature used to train a target machine learning (ML) model. The present invention may also include mapping the determined at least one feature to at least one location of a data store including at least one personal data associated with the determined at least one feature. The present invention may further include retrieving a data record of the at least one personal data associated with the mapped at least one feature from the at least one location of the data store. The present invention may also include determining that the target ML model includes a trace of the retrieved data record. The present invention may further include marking the target ML model as containing the at least one personal data.