Deep Learning Image Classification for Sensitive Document Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current data loss prevention (DLP) technologies face challenges in accurately detecting sensitive information in images, particularly with image-borne identification documents and screenshot-borne data, due to limitations in optical character recognition (OCR) accuracy and the need for large datasets of labeled images, which raises privacy concerns and increases computational costs.

Innovation Solution

The implementation of a deep learning-based image classification system that uses convolutional neural networks (CNNs) to classify images without requiring extensive labeled datasets, by re-training the last few layers of pre-trained models with specialized images and storing extracted features instead of raw images, allowing for continuous refinement and minimizing privacy risks.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If optical character recognition (OCR) is used to detect sensitive information in images, then text extraction capability is provided, but detection accuracy is insufficient and computational costs increase

Engineering Contradiction:
Improvedetection accuracyVSAvoidcomputational costs
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent replaces traditional OCR-based text extraction and classification mechanisms with a deep learning-based image classification system. Convolutional neural networks (CNNs) are trained to directly classify images containing sensitive information without requiring text extraction, thereby substituting the mechanical OCR process with a more efficient neural network-based approach that improves accuracy while reducing computational overhead.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent implements pre-training of deep learning models using large-scale datasets, followed by fine-tuning on specialized sensitive information datasets. This preliminary action of pre-training enables the model to learn general image features before specializing in sensitive information detection, thereby improving detection accuracy while reducing the computational resources needed during actual detection operations.

Inventive Principle:
Principle #10Preliminary action

2Reliability

If large datasets of labeled images are collected for training, then model detection capability improves, but privacy concerns increase and storage requirements increase

Engineering Contradiction:
Improvemodel detection capabilityVSAvoidprivacy concerns
Core Design Contradiction:
ReliabilityVSObject-affected harmful factors

Solution Approach 1:

The patent extracts only the essential training requirements from large labeled datasets by using transfer learning. Pre-trained models on general image datasets are fine-tuned with smaller amounts of specialized sensitive information data, thereby extracting only the necessary training signal while avoiding the need to store and process large volumes of potentially sensitive labeled images.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent uses synthetic image generation and data augmentation techniques to create training data copies and variations. Instead of relying on large collections of real sensitive information images, the system generates synthetic equivalents that preserve the detection learning objectives while eliminating privacy concerns associated with storing actual sensitive data.

Inventive Principle:
Principle #26Copying

3Measurement precision

If deep learning models are trained with specialized images, then detection precision for sensitive information improves, but training time and computational resources increase

Engineering Contradiction:
Improvedetection precisionVSAvoidtraining time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent performs pre-training of deep learning models on large-scale general image datasets before fine-tuning on specialized sensitive information datasets. This preliminary action enables the model to learn general image features and patterns, reducing the training time and computational resources required for the subsequent specialized training phase while maintaining high detection precision.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent implements fine-tuning of only the final layers of pre-trained neural networks rather than retraining the entire model. This partial training approach focuses computational resources on the most critical detection layers, achieving high detection precision for sensitive information while significantly reducing overall training time and resource requirements.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP3920093A1Detecting image-borne identification documents for protecting sensitive information
Publication Date: 2021.12.08 NETSKOPE INC
  • EP3920093A1 patent drawingFigure 1A
  • EP3920093A1 patent drawingFigure 1B
  • EP3920093A1 patent drawingFigure 2

AI summary

Disclosed is detecting identification documents in images (image-borne identification documents) and protecting against loss of the image-borne identification documents - training a DL stack by forward inference and back propagation using labelled ground truth data for the image-borne identification documents and the examples of other image documents. The DL stack includes a first set of layers closer to an input layer and a second set of layers further from the input layer, the first set of layers being pre-trained to perform image recognition before exposing the second layer of the DL stack to the labelled ground truth data for the image-borne identification documents and the examples of other image documents. Also included is storing parameters of the trained DL stack for inference from production images, and using a production DL stack with the stored parameters to classify at least one production image by inference as containing a sensitive image-borne identification document.