On-Premises Image Classifier Using Non-Invertible Feature Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing DLP technologies face challenges in accurately detecting sensitive information in images due to high computational demands of OCR and the need for large-scale labeled data, while customers are hesitant to share their sensitive data with DLP providers due to privacy concerns.
Innovation Solution
A Train Your Own Classifier (TYOC) mechanism using a customized deep learning stack that extracts non-invertible features from images within an on-premises Docker container, allowing organizations to train classifiers without sharing original data, ensuring data privacy and security.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If OCR technology is used to extract text from images for sensitive information detection, then text detection capability is improved, but computational resource demand and processing time increase significantly
Solution Approach 1:
The patent extracts only the essential visual features needed for sensitive information detection without performing full OCR text extraction. The deep learning model directly analyzes image features to identify sensitive content types (passports, driver's licenses, etc.) without converting all text to machine-readable format, thereby reducing computational overhead while maintaining detection accuracy.
Solution Approach 2:
The patent uses deep learning models to create feature representations (embeddings) that capture the essential characteristics of sensitive documents. Instead of processing the entire original image through computationally intensive OCR, the system creates compressed feature vectors that preserve detection capability while requiring minimal computational resources for classification.
2Measurement precision
If deep learning is used to detect images with sensitive information, then detection accuracy is improved, but the need for large-scale labeled training data increases
Solution Approach 1:
The patent pre-trains deep learning models using publicly available datasets containing various document types and formats. This preliminary training establishes a robust foundation for recognizing sensitive document characteristics. When deploying to specific organizations, the pre-trained models require only minimal organization-specific labeled data for fine-tuning, dramatically reducing the total training data volume needed compared to training from scratch.
Solution Approach 2:
The patent employs transfer learning where pre-trained models are adapted to organization-specific requirements by adjusting only the final classification layers. This parameter change approach allows the system to leverage knowledge from large public datasets while requiring minimal organization-specific labeled data, achieving high detection accuracy with small training sets.
3Measurement precision
If organizations share sensitive training data with DLP service providers for model training, then classifier accuracy for organization-specific data is improved, but data privacy and security concerns worsen
Solution Approach 1:
The patent introduces feature embeddings as an intermediary representation between raw sensitive images and the classification model. Organizations extract features from their sensitive images using local deep learning models, and only these non-invertible feature representations are shared with the DLP service provider. The intermediary features preserve classification information while eliminating the ability to reconstruct original sensitive images, thus protecting data privacy.
Solution Approach 2:
The patent creates non-invertible copies of sensitive data in the form of feature embeddings. These embeddings capture essential classification information but cannot be reversed to reconstruct the original sensitive images. Organizations can share these feature copies with service providers for model training without exposing the actual sensitive content, maintaining both accuracy and privacy.
4Device complexity
If traditional DLP methods are used for image analysis, then implementation simplicity is maintained, but detection accuracy for non-ideal images worsens
Solution Approach 1:
The patent applies preliminary image preprocessing operations (normalization, augmentation, resizing) to transform non-ideal images into a standardized format suitable for deep learning analysis. This preliminary action prepares images with variations in quality, orientation, and lighting conditions, enabling the model to achieve high detection accuracy across diverse real-world scenarios while maintaining a relatively simple overall system architecture.
Data Source
Figure 1A
Figure 1B
Figure 2A
AI summary
Disclosed is a method of building a customized deep learning (DL) stack classifier to detect organization sensitive data in images, referred to as image-borne organization sensitive documents, and protecting against loss of the image-borne organization sensitive documents, including distributing a trained feature map extractor stack with stored parameters to an organization, under the organization's control, configured to allow the organization to extract from image-borne organization sensitive documents, feature maps that are used to generate updated DL stacks, without the organization forwarding images of organization-sensitive training examples, and to save non invertible feature maps derived from the images, and ground truth labels for the image. Also included is receiving organization-specific examples including the non-invertible feature maps extracted from the organization-sensitive documents and the ground truth labels and using the received organization-specific examples to generate a customer-specific DL stack classifier. Further included is sending the customer-specific DL stack classifier to the organization.