<p>The widespread presence of fake profiles on social media platforms has emerged as a critical challenge, undermining digital authenticity, user trust, and online security. Addressing this issue, the present study proposes a novel and optimized Transformer-based architecture tailored for the accurate classification of fake and genuine user profiles. Unlike traditional machine learning techniques that depend heavily on manual feature engineering, our approach leverages the deep representational capabilities of Transformer models, which utilize multi-head self-attention mechanisms to capture complex dependencies and subtle patterns within high-dimensional user data. The core of the architecture is built upon an encoder-only Transformer design, incorporating positional encodings to preserve sequential relationships among user behavior metrics, and multi-head scaled dot-product attention layers that allow the model to focus selectively on influential features such as followers_count, fav_number, and statuses_count. The model was trained and evaluated on a real-world dataset. A correlation-based heatmap analysis was performed to identify and prioritize features contributing most to fake profile detection. To further enhance performance, hyperparameters were tuned using the Tree-structured Parzen Estimator (TPE), optimizing dimensions such as the number of attention heads, embedding size, and learning rate. Quantitative results affirm the superiority of the proposed model: it achieved an overall accuracy of 98.74%, precision of 98.89%, recall of 98.55%, and F1-score of 98.72%, markedly outperforming conventional classifiers. The attention mechanism provided valuable interpretability by highlighting which input features were most critical in the classification process, thus enhancing transparency and trust in the model’s decisions.</p>