IO-X: Detecting and Attributing Content-Duplicating Influence Operations on X (Twitter)
Conference
Shafin, AA, Siddique, MN, Carbunar, B. (2026). IO-X: Detecting and Attributing Content-Duplicating Influence Operations on X (Twitter)
. 1809-1818. 10.1145/3774904.3792666
Shafin, AA, Siddique, MN, Carbunar, B. (2026). IO-X: Detecting and Attributing Content-Duplicating Influence Operations on X (Twitter)
. 1809-1818. 10.1145/3774904.3792666
Social platforms' data-access restrictions hinder independent study of influence operations (IOs). We present a formal model of IO undetectability, frame it as a security game and show that content duplication induces detectable gaps. We introduce IO-X, a system for detecting multimodal, content-duplicating IOs (CD-IOs). IO-X links near-duplicate text and images via embedding-based multigraph analysis with sentiment filtering. This uncovers meme-style campaigns missed by text-only methods. It attributes campaigns using ForeignScope, that combines language detection, belief propagation, and timezone inference to flag foreign-linked accounts, and PAN, which integrates profile, content, network, and LLM-derived features for political alignment. Applied to fact-checker-seeded data from the 2024 U.S. and Indian elections, IO-X uncovered 27k+ CD-IOs in the U.S. and 37k+ in India, including 30k+ campaigns duplicating false information, specious sources, or hate speech, and revealing strong partisan asymmetries.
