Inductive detection of influence operations via graph learning.

Area of Science:

• Artificial Intelligence
• Social Media Analysis
• Computational Social Science

Background:

• Influence operations pose a significant threat to public discourse and democratic processes.
• Emerging AI technologies present new challenges for detecting and disrupting large-scale manipulation campaigns.
• Existing detection methods struggle to identify novel operations that evade current security measures.

Purpose of the Study:

• To develop a novel inductive learning framework for detecting influence operations.
• To identify generalizable content- and graph-based indicators of coordinated manipulation.
• To evaluate the framework's ability to detect operations from diverse geopolitical origins.

Main Methods:

• Developed an inductive learning framework incorporating content and graph-based indicators.
• Utilized graph learning to encode abstract signatures of coordinated manipulation.
• Trained and tested detection models across influence operations from Russia, China, and Iran.

Main Results:

• The inductive learning framework demonstrated strong cross-operation generalization capabilities.
• Identified salient, operation-agnostic indicators of influence operations.
• The approach complements existing transductive methodologies, enhancing overall detection coverage.

Conclusions:

• The proposed inductive learning framework offers a generic and effective approach to detecting novel influence operations.
• This method enhances the ability to identify and disrupt large-scale manipulation campaigns, safeguarding public discourse.
• The findings underscore the importance of adaptable AI-driven solutions for combating evolving online threats.