Graph Neural Networks Power ALETHEIA, Advancing Social Media Campaign Detection with 96.6% Accuracy

Influence campaigns present a significant and escalating threat to online platforms, demanding new strategies to protect users from manipulation. Mohammad Hammas Saeed, Isaiah J. King, and Howie Huang, all from George Washington University, address this challenge with ALETHEIA, a novel system designed to both detect malicious accounts and predict their future behaviour on social media. This research demonstrates that analysing the structural characteristics of influence campaigns, combined with linguistic features, substantially improves detection rates compared to traditional methods, achieving a 3.7% increase in F1-score. Crucially, ALETHEIA also introduces a temporal link prediction mechanism, accurately forecasting interactions between malicious accounts and vulnerable users with an impressive average AUC of 96.6%, offering a proactive approach to mitigating the spread of disinformation and coordinated activity online.

The system quantifies the number of accounts used in these operations and forecasts their behaviours within social media networks, achieving an Area Under the ROC Curve (AUC) of 0.81. Researchers analyse influence campaigns from different countries, highlighting that detection pipelines built over a graph-based representation of campaigns, using a mix of topological and linguistic features, offer improvement over standard interaction features.

ALETHEIA utilises state-of-the-art Graph Neural Networks (GNNs) for detecting malicious users, scaling to large networks and achieving a 3.7% F1-score improvement over standard classification methods. The system employs a first temporal link prediction mechanism built for influence campaigns, allowing it to anticipate future malicious activity.

Graph Neural Networks Detect Disinformation Networks

This research investigates the detection and understanding of state-sponsored disinformation campaigns, particularly on platforms like X. The work moves beyond identifying individual troll accounts to understand the network and context of these operations, employing a graph-based approach that leverages contextual and semantic information. Researchers utilise Graph Neural Networks (GNNs) to model relationships between accounts and content, alongside Sentence-BERT (SBERT) for semantic understanding of text and hypergraphs to capture complex relationships. The methods demonstrate improved performance in detecting state-sponsored troll accounts compared to existing techniques, highlighting the importance of analysing the network structure of disinformation campaigns to reveal coordinated behaviour and amplification strategies. Semantic analysis with SBERT helps identify shared narratives, while hypergraphs capture complex relationships missed by traditional graphs. The research is grounded in real-world data, analysing campaigns linked to Russia and other state actors, and acknowledges the potential for unintended consequences of overly aggressive content removal.

Graph Networks Detect Online Influence Campaigns

Scientists have developed ALETHEIA, a novel system designed to detect and forecast the behaviour of malicious accounts involved in online influence campaigns, achieving significant advancements in identifying coordinated activity on platforms like Reddit and X. The work models these campaigns as networks, recognising that troll accounts rely on collaborative efforts and strategic interactions, and frames detection and prediction as node and link detection tasks within a social media graph. Experiments reveal that utilising graph-based approaches improves node classification scores for accounts engaged in these campaigns, demonstrating the importance of structural information. ALETHEIA employs a multi-modal pipeline built upon GraphSAGE, integrating topological features with language embeddings to represent malicious campaigns.

Tests on campaigns from multiple countries demonstrate consistent results, with the system attaining an average F1-score of 96.44% on Reddit and 97.9% on X. The team developed a temporal link prediction mechanism, stacking a Graph Neural Network over a Recurrent Neural Network, to forecast future interactions between troll accounts and regular users, accurately predicting these links with an average Area Under the Curve (AUC) of 96.6%. Specifically, ALETHEIA predicts both troll-to-troll edges and troll-to-user edges, enabling the identification of regular users potentially affected by malicious influence efforts, and offering a powerful tool for preemptive measures against online manipulation.

Forecasting Malicious Influence Campaigns With Aletheia

ALETHEIA represents a major advance in the detection and forecasting of malicious behaviour on online social media platforms, particularly in addressing the growing threat of coordinated influence campaigns. The system analyses data from platforms such as Reddit and X by combining network structure with linguistic features to identify accounts involved in coordinated malicious activity. By leveraging graph neural networks, ALETHEIA improves detection accuracy by 3.7% compared to existing approaches that rely solely on user interaction data.

In addition, ALETHEIA introduces a novel temporal link prediction mechanism that accurately forecasts future interactions between malicious accounts and regular users, achieving an average area under the curve (AUC) of 96.6%. This predictive capability enables the proactive identification of users who may be targeted by harmful influence operations, thereby strengthening online safety and trust. Designed with scalability in mind, ALETHEIA employs efficient graph learning techniques that allow analysis of large, evolving networks using standard computing infrastructure. Researchers stress that its deployment should be complemented by human oversight to ensure responsible use and the protection of free expression.