International Journal of Science, Engineering and TechnologyAuthors: Rishu Kumari, Ridhi Jain, Sanidhya Pal, Nitesh Kaushik, Meenakshi Sharma
Abstract: Online social networks face persistent threats from Sybil attacks in which a single adversary creates many fake identities to manipulate trust ratings, spread misinformation, or claim advertising rewards. This study presents SybilBelief, a semi-supervised framework that treats Sybil detection as probabilistic inference over a Markov random field built from the network graph. The method propagates belief through loopy belief propagation while learning a single homophily parameter θ from a small set of labeled seed nodes through pseudo-likelihood maximization. To prevent numerical underflow during high-degree message products, all message updates are performed in the log domain using the log-sum-exp trick. The edge parameter is optimized via gradient ascent on a pseudo-likelihood objective, implemented through PyTorch’s automatic differentiation applied to an unrolled 50-iteration belief propagation graph, with gradient checkpointing employed to manage memory. Rather than relying on a fixed global threshold for classification, thresholds are tuned per dataset via nested validation, and the Area Under the Precision-Recall Curve (AUPRC) is reported as the primary threshold-invariant metric. Experiments on four datasets — a synthetic Barabási-Albert graph, Epinions, Slashdot Zoo, and a Twitter sample — demonstrate that SybilBelief significantly outperforms SybilRank and SybilLimit under severe label scarcity (0.5% seeds; p < 0.01 in all cases), degrades more gradually under infiltration attacks (macro-F1 remains above 0.7 up to approximately 7–8 infiltration edges per Sybil node), and achieves a crossover with GCN performance at approximately 5% labeled seeds. Targeted seed contamination degrades macro-F1 by approximately twice the margin of equivalent random noise, confirming that adversarial label corruption represents a distinct and more severe threat than random label noise.
DOI: http://doi.org/10.5281/zenodo.21716025
