Content-Agnostic Early Detection of Fake Accounts in Growing Social Networks

The rapid spread of misinformation in online social networks is frequently amplified by fake or automated accounts that manipulate information diffusion and user perception. Detecting such accounts at an early stage remains a significant challenge, as traditional detection methods rely heavily on behavioral or content-based signals that are unavailable for newly created users. This work investigates whether fake accounts can be identified using only structural information derived from network growth dynamics. We propose a content-agnostic framework for early fake account detection based exclusively on topological features extracted from evolving social graphs. Networks are modeled using two preferential attachment mechanisms, namely the Barabási–Albert (BA) and Barabási–Fitness (BF) models, enabling controlled simulation of realistic network evolution. Synthetic graphs containing injected fake nodes are generated, from which interpretable node-level structural features are computed. Multiple supervised machine learning classifiers are then trained to distinguish genuine from fake accounts. Experiments under BA and BF growth show that topology-only features often yield high overall accuracy, yet precision, recall, and F1 for the fake class remain moderate because fake accounts are the minority. Gradient Boosting and Extra Trees provide the most balanced positive-class performance, whereas accuracy-leading models such as Random Forest can under-detect fakes, and KNN can degrade sharply under fitness-driven attachment at high fake prevalence. These findings suggest that early structural signals are informative but incomplete, and that content-agnostic detectors should be assessed with class-sensitive metrics rather than accuracy alone.