Deep Learning Data Augmentation via Anonymized Feature Similarity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Insufficient training data for deep learning models, particularly in the business sector where customer data privacy concerns limit data sharing, leads to difficulties in achieving accurate output results and increases the risk of information leakage, especially when personal information is present in intermediate feature values.
Innovation Solution
A learning method that augments original training data by adding dummy information similar to non-stored target information, such as license plates or face images, to generate additional training data sets, which are then processed to ensure intermediate feature values resemble reference feature values without personal information, thereby increasing the usable data for model learning without storing sensitive information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If data augmentation is executed on original training data to expand the training data range, then the amount of training data increases, but the risk of information leakage increases when personal information is present in the data
Solution Approach 1:
The patent extracts and removes personal information (such as face images, license plate numbers, and other identifiable data) from the training data before performing data augmentation. This extraction process eliminates the harmful elements while preserving the useful features for model training, thereby increasing training data quantity without proportionally increasing information leakage risk.
Solution Approach 2:
The patent introduces an intermediary processing step that anonymizes or pseudonymizes personal information in the training data. This intermediary transformation maintains the structural and feature characteristics needed for learning while replacing sensitive information with non-identifiable representations, allowing data augmentation to proceed safely.
2Quantity of substance
If intermediate feature values are stored to increase usable data for sequential learning, then the amount of training data increases, but personal information may be retained in these feature values creating security risks
Solution Approach 1:
The patent converts the potentially harmful intermediate feature values that contain personal information into beneficial anonymized feature representations. By processing these feature values through anonymization techniques, the system transforms data that could leak personal information into safe, privacy-preserving representations that maintain learning utility for sequential training of deep learning models.
3Quantity of substance
If customer data is permanently stored or shared for different customers' tasks, then the amount of training data increases, but contract obligations and information leakage risks increase
Solution Approach 1:
The patent creates anonymized copies of customer data that can be freely shared and used for training different deep learning models without violating contracts or security requirements. These copied, anonymized datasets retain the essential features and patterns needed for training while eliminating personally identifiable information, enabling data reuse across multiple customers and applications.
Data Source
AI summary
A learning method executed by a computer, the learning method including augmenting original training data based on non-stored target information included in the original training data to generate a plurality of augmented training data, generating a plurality of intermediate feature values by inputting the plurality of augmented training data to a learning model, and learning a parameter of the learning model such that, with regard to the plurality of intermediate feature values, each of the plurality of intermediate feature values generated from a plurality of augmented training data, augmented from reference training data, becomes similar to a reference feature value.