Token Synthesis Using Perturbed Vectors for Sensitive Data Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models struggle to accurately identify sensitive information due to evolving definitions and varied formats, leading to false indications and inadequate training sets, which are exacerbated by low frequency of non-standard data.
Innovation Solution
Synthesizing additional tokens based on erroneous associations using autoencoders and clustering criteria to retrain the neural network, enhancing its ability to detect sensitive data in various forms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a rule-based system with regular expression patterns is used to detect sensitive information, then the system can capture standard formats of sensitive data, but it fails to accurately capture non-standard or varied formats of sensitive information
Solution Approach 1:
The patent transforms the detection approach from fixed rule-based parameters to dynamic machine learning model parameters. The system uses training data to learn patterns of sensitive information, allowing the model to adapt to various formats including non-standard ones. The model's parameters are updated through training with positive examples (actual sensitive data) and negative examples (non-sensitive data), enabling it to generalize across different formats without requiring explicit rules for each format variation.
2Reliability
If a machine learning model is trained on existing data to identify sensitive information, then the model can learn patterns from training examples, but the low frequency of non-standard sensitive data in training sets results in inaccurate detection and false indications
Solution Approach 1:
The patent applies data augmentation techniques as a preliminary action before model training. Synthetic examples of sensitive information are generated using autoencoders and perturbation methods, creating expanded training datasets that include rare and non-standard formats. This preliminary data preparation ensures that the model encounters diverse examples during training, improving its ability to reliably detect sensitive information across various formats without requiring vast amounts of real-world training data.
3Adaptability or versatility
If the training set is expanded to include more variations of sensitive data, then the model's ability to detect diverse formats improves, but the complexity of preparing and managing the training set increases
Solution Approach 1:
The patent implements self-service through automated synthetic data generation. The system uses autoencoder models to automatically generate synthetic sensitive information examples by encoding real examples and decoding them with perturbations. This self-generating capability eliminates the need for manual creation and management of diverse training examples, reducing the complexity of training set preparation while expanding format coverage. The system serves itself by automatically creating the training data it needs.
4Adaptability or versatility
If a neural network is used to detect sensitive information, then the model can handle varied formats better than rule-based systems, but erroneous associations occur when the model fails to correctly categorize tokens
Solution Approach 1:
The patent implements feedback mechanisms through iterative training and evaluation. The system trains the model on synthetic and real examples, evaluates its performance, identifies erroneous associations, and uses these errors to generate additional targeted synthetic examples for retraining. This feedback loop continuously improves categorization accuracy by addressing specific weaknesses in the model's performance, allowing it to handle varied formats more reliably over time.
Data Source
AI summary
A method includes generating a first vector associated with a token using a first set of parameters of a first learning model based on the token, determining a prediction indicating that the token is associated with a first label based on a set of clustering criteria, the first vector, vectors of a first vector set, and vectors of a second vector. The method includes generating a perturbed vector associated with a second label by modifying a value of the first vector and updating the second vector set to comprise the perturbed vector. The method also includes generating a synthesized token associated with the second vector set based on the perturbed vector using a second set of parameters of the first learning model and training a second learning model based on the synthesized token.