RB-GNN: Securing Social Collaboration via Residual Behavior Graph Learning
Year: 2026
Type: conference-paper
Abstract: Online social platforms have evolved into vital ecosystems for mass collaboration and information sharing. However, the integrity of these cooperative environments is increasingly compromised by sophisticated malicious accounts (e.g., social bots) that disrupt public trust. While Graph Neural Networks (GNNs) have shown promise in detection, most existing methods rely heavily on explicit social connections (e.g., follower/following links). These explicit structures are often sparse or deliberately manipulated by adversaries to camouflage within legitimate communities, leading to detection failures. To address this, we propose a novel framework, RB-GNN (Residual Behavior Graph Neural Network), designed to secure social collaboration by uncovering latent behavioral patterns. First, we introduce an Operation-Conditioned Behavior Encoder that contextualizes textual semantics with user action types to capture precise operational intent. Crucially, we construct a Behavior-Residual Graph to mine implicit cooperative affinities that are absent from the observed topology, effectively bypassing structural camouflage. Finally, a topology-aware gated fusion mechanism is employed to adaptively integrate explicit structural signals with implicit behavioral evidence. Extensive experiments on public benchmarks (Cresci-2015 and TwiBot-20) demonstrate that our approach consistently outperforms state-of-the-art baselines, offering a robust defense for trustworthy online collaboration.
Authors: Zhenyuan Wang, Lixin Ji, Shuxin Liu
Institution: PLA Information Engineering University
Language: English
Topic: Spam and Phishing Detection
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7167487015