Synthetic Email Text Detection Using Sender and Chunk Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large language models are vulnerable to malicious use in cyberattacks, particularly in email-borne attacks, and existing detection techniques for synthetic text are unreliable, especially on short texts, posing a significant risk to organizations.
Innovation Solution
A novel synthetic email detection technology that analyzes the 'From' header of emails to identify the sender as an individual, extracts textual content, splits it into parts, and uses multiple synthetic text detection engines to detect and flag emails containing synthetic text generated by large language models, employing methods like watermarking and binary classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing detection techniques for synthetic text are used, then detection capability is provided, but reliability is poor especially on short texts
Solution Approach 1:
The email text is segmented into multiple chunks or segments for analysis. The detection system processes different portions of the email separately and aggregates results, which improves detection reliability especially for short texts where individual segments can still be analyzed effectively
Solution Approach 2:
The detection approach moves from analyzing only the email body text to incorporating metadata dimensions such as the 'From' header analysis. By examining sender information, email structure, and metadata alongside text content, the system achieves more reliable detection without being constrained by text length
2Measurement precision
If multiple synthetic text detection engines are used, then detection accuracy is improved, but device complexity increases
Solution Approach 1:
Multiple synthetic text detection engines are merged into a unified detection system that processes emails through several analysis stages. The system combines results from multiple detection models and integrates them with metadata analysis, achieving higher accuracy while managing complexity through structured integration
Solution Approach 2:
An intermediary processing layer is introduced that coordinates multiple detection engines. This intermediary manages the workflow between different detection models, aggregates their outputs, and presents a unified detection result, reducing the perceived complexity for users while maintaining high detection accuracy
3Reliability
If email filtering capabilities are enhanced, then security against cyberattacks is improved, but processing time increases
Solution Approach 1:
The system performs preliminary analysis on email metadata such as the 'From' header before conducting full text analysis. By pre-processing and filtering based on metadata indicators, the system quickly identifies suspicious emails that need detailed analysis while dismissing legitimate emails faster, reducing overall processing time while maintaining security reliability
Data Source
AI summary
A computer-implemented method of detecting synthetic text in an email may determine whether the sender of a received email is a named individual and if so, may extract textual content from the email. When the size of the extracted textual content is at least equal to a configurable size threshold, the extracted textual content may be analyzed using a plurality of synthetic text detection engines, each being configured to generate a synthetic text detection score. The synthetic text detection scores may then be compared to a respective score threshold associated with each of the synthetic text detection engines. An action to warn and/or protect the recipient of the email may be carried out when one or more synthetic text detection scores is at least equal to a respective associated score threshold, which is indicative of a detection of synthetic textual content in the received email.