The increasing spread of spurious profiles as well as bot networks on social media call for detection methods that are advanced enough to detect even sophisticated fake profiles. Approaches so far have focused on static features or simple behavioral models with little to no reflection on the complexity and dynamic nature of the fake profiles and, therefore translate to low detection accuracy and delayed identification of suspicious accounts. We therefore propose a robust fake profile detection model that uses hybrid deep learning techniques, multimodal analysis, and temporal dynamics to address those challenges. Our proposed system employs a combination of the CNN-Transformer hybrid model for comprehensive feature extraction from users’ interaction sequences while the static profile traits such as profile pictures, bio texts are extracted by CNNs, and the long-term dependencies of user behavior are captured using Transformers. We incorporate a Multimodal Attention Network (e.g., M-BERT) to combine three kinds of input data: textual profile information, images, and behavioral data. The model could thus recognize anomalies across all dimensions. To integrate the changes in behavior patterns over time, we implemented the use of TCNs to identify emergent patterns, such as sudden spikes in followers or an irregular rate of engagement-often easily indicative of bot behavior. This hybrid architecture has shown crucial performance improvement, with an accuracy of 94–96%, along with precision and recall of 0.90–0.93. The multimodal and temporal analysis allow the dynamic real-time assessment of authenticity profiles with a reduction of detection delay down to 20–25%. The proposed model advances the state-of-the-art in detecting fake profiles through proper capture of static and dynamic features, eliminating inconsistency across multimodal data streams, thereby achieving more accurate and timely detection operations.