Parallel Deep Learning for Cybersecurity-Oriented Multi-Platform Social Media Bot Detection

Abstract

Social media is a key unit of today's cyber communication society; however, being widely available, social media networks also contribute to the diffusion of social bots, which are then abused for forging global digital weapons, such as spreading fake news, organized manipulation and infiltration, or malicious propaganda actions. However, identifying these bots in a multi-platform setting is difficult due to the heterogeneity of data structures and behaviours, as well as platform-specific issues. In this paper, we have presented a parallel DNN-based social media bot detection model for various platforms. The proposed architecture has a multi-head parallel structure sharing the latent representations and simultaneously learning the shared latent features and platform-specific independent detections. What it allows, however, is for the framework to handle both generic bot prototype behaviours and platform-specific functionalities in a single solution that scales well. The proposed approach is evaluated on two real-world datasets taken from Twitter and TikTok, including behavioural, engagement, and account-level features. Experiments show strong and consistent detection performance across the tested platforms. The detection heads on Twitter reach 94.47% accuracy, 94.42% F1-score, and a ROC-AUC of 94.77%, while those on TikTok obtain an accuracy of 91.29%, an F1-score of 92.92% and a ROC-AUC score of 96.30%. High precision–recall performance also confirms the effectiveness of the proposed system in detecting automated harmful accounts with low false positive rates.