Context-Aware Chat DLP Using Sliding Windows and Language Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data loss prevention (DLP) systems for Software-as-a-Service (SaaS) applications lack contextual knowledge to accurately detect sensitive data in chat messages, often misidentifying non-sensitive alphanumeric strings as sensitive data.
Innovation Solution
A context-based sensitive chat message DLP system that maintains a sliding window of messages within chat contexts, using an entropy-based filter and a DLP language model with a sentence transformer to classify messages as sensitive or non-sensitive, adapting to the conversation context and user interface format.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional DLP systems analyze chat messages without context, then detection speed is maintained, but false positive rate increases due to lack of contextual knowledge
Solution Approach 1:
The system pre-processes and stores chat history context in vector embeddings before analysis occurs. When a message needs detection, the relevant contextual vectors are already prepared and can be quickly retrieved and compared, eliminating the need for complex real-time contextual analysis while maintaining high detection accuracy.
Solution Approach 2:
The patent introduces vector embeddings as an intermediary representation layer between the raw chat messages and the detection engine. This intermediary transforms unstructured text into structured numerical vectors that capture semantic meaning, enabling accurate contextual comparison without requiring the detection system to directly parse and understand natural language context.
2Measurement precision
If contextual analysis is added to DLP systems, then false positives are reduced, but processing time increases
Solution Approach 1:
Chat messages are pre-processed into vector embeddings and stored in a searchable index before they need to be analyzed for sensitive data. This preliminary transformation allows the detection system to perform fast vector similarity searches rather than expensive natural language processing during actual detection, significantly reducing processing time while maintaining contextual accuracy.
Solution Approach 2:
The system changes the representation parameters of chat context from raw text to vector embeddings. This parameter transformation enables efficient mathematical operations and similarity comparisons that are computationally much faster than traditional text-based contextual analysis, thereby reducing processing time while preserving contextual information.
3Measurement precision
If sliding window of chat history is maintained, then contextual understanding improves, but memory consumption increases
Solution Approach 1:
The system extracts only the essential contextual information from chat history and stores it as compact vector embeddings, rather than storing the complete chat history text. This extraction process retains the semantic meaning needed for contextual understanding while dramatically reducing the storage requirements, as vectors occupy far less memory than raw text.
Solution Approach 2:
By transforming chat history into vector representations, the system changes the data structure from text-based to numerical-based storage. This parameter change enables more efficient memory utilization, as vectors can be stored in compact binary formats and processed in batches, reducing overall memory consumption while preserving contextual semantics.
Data Source
AI summary
A context-based chat message data loss prevention system (“DLP system”) detects sensitive chat messages communicated via Software-as-a-Service (“SaaS”) applications. The DLP system receives chat messages via SaaS connectors and buffers the chat messages in sliding windows that correspond to context of chat messages within UIs of the SaaS application. The DLP system then filters messages in the sliding windows and classifies the filtered messages with a language model. The resulting sensitive/non-sensitive classifications by the language model thus incorporate chat context for corresponding SaaS applications.