<![CDATA[Facial expression recognition (FER) is a relevant field of study with applications in human-computer interaction, healthcare, and security. Although recent approaches demonstrate excellent outcomes on the recognition of basic emotions, the authenticity of expressions (genuine versus fake) remains unexplored. In this work, we propose a dual-task deep learning framework based on EfficientNet-B0, enhanced with a lightweight squeeze-and-excitation (SE) attention mechanism, to collaboratively work on multiclass emotion recognition (seven categories: angry, disgust, fear, happy, neutral, sad and surprise) and authenticity classification (genuine vs fake). The architecture leverages a shared backbone for representing feature, followed by task-dedicated branches trained using categorical cross-entropy and focal loss, respectively. To overcome the lack of publicly available benchmarks incorporating authenticity labels, we designed a curated dataset annotated with both emotional categories and authenticity information. Experimental evaluation demonstrates that the proposed dual-task model with the SE attention mechanism achieves 98.5% accuracy for emotion recognition and 92.2% accuracy for authenticity prediction, emphasizing both the effectiveness of the framework and the inherent challenges of authenticity detection. Moreover, we present a deployable real-time system demonstrating the feasibility of integrating authenticity-aware FER into practical applications such as e-learning analytics, security surveillance, and affective computing.]]>
