Multi-Modal Embeddings for Cross-Platform Coordinated Information Campaigns
Authors: Fabio Barbero, Sander op den Camp, Kristian van Kuijk et al.
Published: 2023-09-22
arXiv: 2309.12764
Source: arXiv (cs.SI)
Abstract
Coordinated multi-platform information operations are implemented in a variety of contexts on social media, including state-run disinformation campaigns, marketing strategies, and social activism. This paper presents a multi-modal approach that identifies social media messages potentially engaged in a coordinated information campaign across multiple platforms. Our approach incorporates textual content, temporal information and the underlying network of user and messages to identify groups with unusual coordination patterns across multiple platforms. Applied to content on four platforms related to the Syrian Civil Defence (White Helmets) — Twitter, Facebook, Reddit, and YouTube — results show that our approach identifies social media posts that link to news YouTube channels with similar factuality score, an indication of coordinated operations.
Why This Work Matters
The critical operational gap in detecting coordinated IO campaigns is cross-platform attribution: a campaign’s signal is distributed across multiple platforms and is invisible when each platform is analyzed in isolation. This paper directly addresses that gap with a multi-modal methodology that fuses text, timing, and network structure — exactly the combination needed to detect coordinated inauthentic behavior (CIB) at operational scale without requiring pre-labeled training data.
The White Helmets case study is analytically significant: the Syrian Civil Defence was simultaneously the subject of pro-Syrian government disinformation (claiming staged propaganda) and genuine civil society support. A methodology that detects the coordination pattern without pre-labeling which side is running the operation has direct counter-IO application.
Core Concepts and Contributions
Multi-modal signal fusion: Three signal types combined — (1) textual embedding (semantic content similarity), (2) temporal patterns (synchronized posting behavior), (3) network structure (user-message interaction graphs). Each signal alone produces false positives; the combination provides discriminative precision.
Cross-platform scope: Applied across four platforms simultaneously. This is the key advance over single-platform CIB detection. A coordinated campaign using different accounts and formats on different platforms remains detectable through temporal and network pattern convergence even when content diverges.
Ground-truth-free operation: Does not require labeled training data of known IO campaigns. Detects coordination patterns statistically anomalous relative to organic behavior, enabling application to novel campaigns — directly addressing the IOHunter companion challenge (IOHunter handles driver detection; this paper handles cross-platform coordination detection).
Connections
- IOHunter: Graph Foundation Model — complementary approach at account/driver level vs. this paper’s message-coordination level
- Social Media Information Operations — Zaman & Chen (2025) — formal IO optimization framework
- Information Operations — primary subject domain
- Disinformation Campaign — the operational activity being detected