A Multimodal Graph-Attentive Framework for Robust Bot and Fake-Account Detection in Online Social Networks

The proliferation of automated and fraudulent accounts on online social networks (OSNs) threatens information integrity, user trust, and platform security by enabling misinformation campaigns, coordinated opinion manipulation, and large-scale fraud. Existing detection approaches typically rely on a single information modality profile metadata, behavioral statistics, textual content, or network topology making them vulnerable to increasingly sophisticated bots that mimic human behavior in unmonitored dimensions. This paper proposes a unified multimodal detection framework, MGA-Bot (Multimodal Graph-Attentive Bot Detector), which jointly encodes four complementary feature families: static profile attributes, temporal behavioral sequences, transformer-based semantic representations of user-generated content, and relational embeddings derived from a heterogeneous graph neural network over the social graph. A cross-modal attention mechanism fuses modality-specific embeddings into a unified account representation, while an adversarial training component improves robustness against evasive human-mimicking bots. In addition, a Shapley-value-based explainability layer supports transparent and auditable moderation decisions. As no empirical experiments have yet been conducted, the contribution of this paper is a theoretically grounded architecture specified through per-module equations, a training objective, and computational-complexity analysis, together with a detailed evaluation protocol and qualitative comparison with state-of-the-art approaches, providing a foundation for future empirical validation.

کلیدواژه ها:

social network security ، bot detection ، fake account detection ، graph neural networks ، natural language processing