Deep Learning Stack for Image-Borne ID Detection
Find Innovative SolutionsGenerate 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 images, due to high computational demands and low accuracy of optical character recognition (OCR) and the need for large datasets of labeled images for deep learning, which raises privacy concerns and is difficult to acquire.
Innovation Solution
The implementation of a deep learning-based image classification system using convolutional neural networks (CNNs) that re-trains models with a small number of specialized labeled images, storing extracted features instead of raw images to maintain privacy and continuously refine accuracy without requiring large upfront datasets, and utilizing machine learning for detecting and enforcing policies on sensitive image content within the Netskope cloud access security broker (N-CASB).
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If optical character recognition (OCR) is used to extract text from images for detecting sensitive information, then text detection capability is provided, but computational resources are excessively consumed and accuracy is insufficient
Solution Approach 1:
The patent replaces the mechanical OCR text extraction process with a deep learning-based image classification system. Instead of using OCR to extract text and then analyze it, the system directly classifies images containing sensitive information using convolutional neural networks, thereby eliminating the high computational overhead and accuracy limitations of OCR while maintaining detection capability.
Solution Approach 2:
The patent changes the detection parameters by training deep learning models on small, specialized datasets of labeled images containing sensitive information. This approach transforms the detection methodology from text-based OCR analysis to image-based deep learning classification, achieving both higher accuracy and reduced computational resource consumption.
2Measurement precision
If deep learning models are trained with large datasets of labeled images to improve detection accuracy, then detection performance is enhanced, but privacy concerns arise and data acquisition becomes difficult
Solution Approach 1:
The patent applies partial training action by using a small, carefully selected dataset of labeled images for training deep learning models, rather than requiring large volumes of sensitive data. This partial approach achieves sufficient detection accuracy while minimizing privacy risks and data acquisition difficulties.
Solution Approach 2:
The patent introduces an intermediary approach where deep learning models are trained on synthetic or publicly available labeled images that represent sensitive information patterns, rather than directly using real sensitive images. This intermediary training data allows the model to learn detection capabilities without compromising actual sensitive information or raising privacy concerns.
3Measurement precision
If deep learning models are trained with large datasets of labeled images to improve detection accuracy, then detection performance is enhanced, but time and cost increase significantly
Solution Approach 1:
The patent uses partial training with a small, targeted dataset of labeled images, eliminating the need for extensive large-scale data collection and processing. This partial approach reduces training time and computational costs while achieving sufficient detection accuracy for practical deployment.
Solution Approach 2:
The patent performs preliminary action by pre-training deep learning models on small, specialized datasets before deployment. This preliminary training establishes baseline detection capabilities that can be continuously refined with production-generated labels, avoiding the need for extensive initial data collection and reducing overall time and cost investment.
Data Source
AI summary
Disclosed is detecting identification documents in image-borne identification documents and protecting against loss of the image-borne identification documents. A trained deep learning (DL) stack is used to classify production images by inference as containing a sensitive image-borne identification document, with the trained stack configured with parameters determined using labelled ground truth data for the identification documents and examples of other image documents. The trained DL stack is configured to include a first set of layers closer to an input layer and a second set of layers further from the input layer, with the first set pre-trained to perform image recognition before exposing the second set of layers of the stack to the labelled ground truth data for the image-borne identification documents and examples of other image documents, and using the inferred classification of the sensitive image-borne identification document in a DLP system to protect against loss by image exfiltration.