Multimodal Fingerprinting Pipelines for Unstructured Data Loss Prevention
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional DLP solutions fail to effectively monitor and protect a diverse range of digital assets, particularly unstructured data formats like images, videos, and audio files, leading to vulnerabilities in data exposure and unauthorized access.
Innovation Solution
A system that generates multimodal fingerprints using AI to classify and protect digital assets by converting files into irreversible representational objects, indexing them in a database, and comparing fingerprints in real-time to detect and block unauthorized data transfers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional DLP schemes are used, then simple and straightforward implementation is achieved, but coverage of certain file types is insufficient leading to data exposure
Solution Approach 1:
The system segments the DLP functionality into multiple specialized components: source connectors for different data sources, input classifiers for file type identification, object extraction pipelines for various file formats, and fingerprinting modules for different data types. Each component handles specific file types independently, enabling comprehensive coverage without requiring a monolithic complex system.
Solution Approach 2:
The system implements universal handlers that can process multiple file types through a common architecture. The object extraction pipelines and fingerprinting modules are designed to work with diverse file formats (images, videos, audio, documents, code) using standardized interfaces, allowing the system to adapt to new file types without fundamental redesign.
2Reliability
If multimodal fingerprinting is implemented, then comprehensive protection across diverse file types is achieved, but processing complexity increases
Solution Approach 1:
The system performs preliminary classification of files by type and format before processing, using input classifiers to route files to appropriate extraction pipelines. This preliminary action organizes the complex processing workload and enables parallel processing of different file types, reducing overall processing complexity while maintaining comprehensive protection.
Solution Approach 2:
The system introduces intermediary components including object extraction pipelines that convert various file formats into standardized representational objects, and fingerprinting modules that transform these objects into comparable fingerprints. These intermediaries simplify the comparison process and enable reliable detection across diverse file types without requiring direct complex analysis of each format.
3Measurement precision
If traditional DLP methods are used, then low processing overhead is maintained, but false positives occur due to inability to distinguish proprietary from public data
Solution Approach 1:
The system changes the parameter of data representation by converting files into multimodal fingerprints that capture essential characteristics in a standardized format. This transformation enables precise comparison and distinction between proprietary and public data while maintaining processing efficiency through optimized fingerprint matching algorithms rather than full-file analysis.
4Loss of information
If comprehensive fingerprinting of all file types is performed, then complete data visibility is achieved, but storage and processing resources increase
Solution Approach 1:
Instead of storing complete copies of all proprietary files, the system creates and stores condensed fingerprint representations that capture the essential characteristics of each file. These fingerprints serve as efficient proxies that enable complete data visibility and comparison while consuming minimal storage resources compared to storing actual file contents.
Data Source
AI summary
Some aspects of the present disclosure relate to a system that uses source connectors to integrate with various user-defined data sources containing proprietary information. The source connectors read data files from the user-defined data sources and send the data files to an input classifier, which associates each data file with a corresponding object extraction pipeline based on a file type and/or encoding format of the data file. The system uses multiple object extraction pipelines to extract representational data objects from the data files. The representational data objects are then processed by various fingerprinting modules of the system. These fingerprinting modules are configured to generate a set of multimodal fingerprints for each data file based on the representational data object(s) associated with the data file. The multimodal fingerprints are then indexed/stored in a database for later use.