Inline Nested DLP Classification with Selective Model Layers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional DLP systems face challenges in handling diverse and complex data types due to overconfidence in predictions, insufficient labeled data, and high computational costs, leading to inaccurate and inefficient data protection.
Innovation Solution
The system employs hierarchical classification using Bert-tiny tokenization, selective sub-layer activation, synthetic data generation, and automated data labeling pipelines powered by Large Language Models (LLMs) to enhance data protection, with features like label smoothing and softmax normalization for calibrated predictions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If traditional rule-based DLP systems are used, then implementation is simple, but accuracy in handling diverse data types deteriorates
Solution Approach 1:
The patent transforms the DLP system from rule-based to machine learning-based, changing the fundamental parameter of classification methodology. This enables the system to handle diverse data types (text, images, PDFs) with high accuracy by learning patterns from training data rather than relying on predefined rules, directly resolving the contradiction between implementation simplicity and classification accuracy.
Solution Approach 2:
The patent replaces the mechanical rule-based classification system with an intelligent machine learning system using BERT models. This substitution allows the system to automatically adapt to diverse data formats and nested hierarchies, achieving high classification accuracy without manual rule configuration for each data type.
2Reliability
If comprehensive data classification is performed, then protection coverage is improved, but computational cost increases
Solution Approach 1:
The patent segments the classification process into distinct hierarchical levels (super-category, sub-category, granular category) and processes only relevant segments based on detected data type. This segmentation allows comprehensive protection coverage while reducing computational cost by avoiding unnecessary classification steps for each data item.
Solution Approach 2:
The patent performs preliminary classification at the super-category level first, then selectively activates sub-layers for further classification only when needed. This preliminary action approach ensures comprehensive protection by establishing baseline classification while reducing computational cost through selective deep processing of only necessary cases.
3Measurement precision
If hierarchical classification is implemented, then classification precision is improved, but processing time increases
Solution Approach 1:
The patent implements a dynamic hierarchical classification system where the depth of classification is adjusted based on the detected data type and context. For simple cases, only super-category classification is performed; for complex cases requiring nested classification, the system dynamically activates additional classification layers, thereby achieving high precision when needed while minimizing processing time for routine cases.
Solution Approach 2:
The patent performs preliminary super-category classification first, then selectively activates sub-layers for deeper classification only when the preliminary classification indicates it is necessary. This preliminary action approach ensures high classification precision for complex cases while reducing processing time by avoiding unnecessary deep classification for simple cases.
4Measurement precision
If more training data is used, then model accuracy is improved, but data preparation complexity increases
Solution Approach 1:
The patent implements automated data labeling pipelines that use pre-trained LLMs to automatically generate labels for training data, eliminating the need for manual annotation. This self-service approach enables the system to utilize large volumes of training data to improve model accuracy while keeping data preparation complexity low, as the automated pipelines handle the complex task of labeling at scale.
Data Source
AI summary
The disclosure presents systems and methods for hierarchical classification of input data across a plurality of categories. A machine learning model processes various data formats, starting with dimensional reduction using tokenization techniques, such as Bert-tiny tokenization, to create model-readable representations. The system predicts super-categories, sub-categories, and granular categories through selective activation of sub-layers tied to identified super-categories, optimizing computational efficiency. Label smoothing during training mitigates overconfidence in predictions, while softmax normalization refines inference outputs. Synthetic data generation using Large Language Models (LLMs) supplements training datasets, and an automated data labeling pipeline efficiently generates hierarchical labels. Modifications to the model, such as stop word removal and file size limitations, further reduce latency. Inference analyzes logits to predict hierarchical paths, providing detailed classifications with clear outputs. The method is adaptable for multimodal formats, ensuring scalable and accurate predictions across diverse data types while minimizing computational costs and improving reliability.