Heterogenous Social Media Analysis for Efficient Deep Learning Fake-Profile Identification

Abstract

The surge in e-scams attributed to an estimated 30% of fake social media accounts has highlighted the urgent need to identify such fraudulent profiles. Since the current model cannot handle multi-model networks, an attempt has been made to solve the real-time problems. This study introduced a cutting-edge deep-transfer learning model that streamlines fake-profile detection through a comprehensive analysis of diverse social media data samples. Our model gathers a wide range of data from various social media platforms, such as posts, likes, comments, multimedia content, user activity, login behaviors, etc. Each data type is individually processed to detect suspicious patterns synonymous with fake accounts - for instance, discrepancies like male profiles predominantly posting about or using images of females. Similarly, audio signals undergo 1D Fourier, Cosine, Convolutional, Gabor, and Wavelet Transforms. In contrast, image and video data are processed with their 2D counterparts. Text data is transformed using Word2Vec, aiding our binary Convolutional Neural Network (bCNN) to distinguish between genuine and fake profiles. Feature optimization is handled by the Grey Wolf Optimizer (GWO) for 2D data and the Elephant Herding Optimizer (EHO) for 1D data, ensuring minimal feature redundancy. Separate 1D CNN classifiers, then classify the refined features to pinpoint fake profiles. The results from these classifiers are amalgamated through a boosting mechanism. Our results reveal an 8.3% increase in accuracy, 5.9% in precision, and 6.5% in recall compared to conventional methods. Testing on initial login and signup processes further validated our model's effectiveness, achieving a remarkable 93.5% accuracy in detecting illegitimate account activities, effectively mitigating cold-start challenges.