The rapid advancement of artificial intelligence-based image generation technologies and generative models has made the detection of fake and manipulated images a major challenge in the fields of computer vision and digital forensics. Images generated by generative models and images manually manipulated by humans can exhibit substantial visual similarity to authentic images, while traditional detection approaches based on human i  ection or individual features often face significant limitations. Therefore, the development of intelligent deep learning-based systems capable of exploiting complementary information contained in images is of considerable importance. In this study, a hybrid multimodal architecture, termed HMFT, is proposed for detecting real and fake face images. The proposed architecture consists of three complementary processing streams: a visual feature stream based on ResNet-50, a frequency and noise feature stream based on the Discrete Cosine Transform (DCT) and Spatial Rich Model (SRM) filters, and a semantic visual feature stream based on the CLIP model. The features extracted from these streams are first projected into a shared feature space and subsequently integrated using a Transformer Fusion module before being provided to the >The proposed model was evaluated on the CIPLAB Real and Fake Face Detection and Hard Dataset (Fake-Vs-Real-Faces), the latter containing GAN-generated fake face images. To improve the reliability of the evaluation, stratified five-fold cross-validation was employed. On the CIPLAB dataset, the final HMFT–HOG–RBF-SVM pipeline achieved a mean accuracy of 99.76%, balanced accuracy of 99.75%, precision of 99.79%, recall of 99.69%, F1-score of 99.74%, ROC-AUC of 99.996%, and PR-AUC of 99.995%. The HMFT–MLP pipeline was also evaluated for comparison and achieved a mean accuracy of 99.51%. On the Hard Dataset, the HMFT–MLP pipeline achieved a mean accuracy of 99.38% and an F1-score of 99.43%, demonstrating the capability of the proposed architecture to detect GAN-generated samples in this dataset and maintain high performance under more challenging conditions. The ablation study further demonstrated that all components of the proposed architecture contribute meaningfully to the final performance. In particular, removing the semantic features extracted by CLIP resulted in a substantial performance degradation, confirming the importance of jointly exploiting visual, frequency, and semantic information. Furthermore, removing the frequency components, visual features, or Transformer Fusion module led to noticeable reductions in the evaluation metrics. Overall, the findings demonstrate that the HMFT architecture, through multi-level feature integration and a Transformer-based attention mechanism, can effectively distinguish between real and fake face images with high accuracy and stability. The proposed architecture can serve as an effective framework for developing intelligent fake-content detection systems in security, media, and digital forensics applications. Keywords: Deep Learning, Fake Image Detection, Digital Forensics, ResNet-50, CLIP, Multimodal Feature Fusion   

   In this study, a novel strategy was devised to evaluate the mechanical properties of semi-crystalline polymer nanocomposites. To this end, the Halpin-Tsai model was comprehensively improved to also include the impact of dispersion quality, nanoparticle aggregation/agglomeration, formation of the polymer/particle interphase region and induced/self-crystallinity. Besides enhancing the tensile modulus of the system, a particular fraction of either dispersed or clustered nanoparticle domains was considered to participate in the induced crystallization process (active domains). The content of active domains was estimated using the Gutzow-Dobreva theory or a specifically designed method based on the requirements of the system. On the contrary, the so-called inactive domains were assumed to remain distributed within the polymer matrix and contribute to the system response mechanism. The mechanical features of the interphase region were defined using De Gennes’s self-similar carpet theory considering the molecular characteristics of the polymer phase. Moreover, the porosity index and bonding nature between aggregated or agglomerated nanoparticles were used to estimate the tensile modulus of the clustered doamins. The model validation process was conducted via benchmarking its predictions against the obtained data from the literature, which showed acceptable agreement and proved the model to be a quite accurate interpretation tool. In this study, a novel strategy was devised to evaluate the mechanical properties of semi-crystalline polymer nanocomposites. To this end, the Halpin-Tsai model was comprehensively improved to also include the impact of dispersion quality, nanoparticle aggregation/agglomeration, formation of the polymer/particle interphase region and induced/self-crystallinity. Besides enhancing the tensile modulus of the system, a particular fraction of either dispersed or clustered nanoparticle domains was considered to participate in the induced crystallization process (active domains). The content of active domains was estimated using the Gutzow-Dobreva theory or a specifically designed method based on the requirements of the system. On the contrary, the so-called inactive domains were assumed to remain distributed within the polymer matrix and contribute to the system response mechanism. The mechanical features of the interphase region were defined using De Gennes’s self-similar carpet theory considering the molecular characteristics of the polymer phase. Moreover, the porosity index and bonding nature between aggregated or agglomerated nanoparticles were used to estimate the tensile modulus of the clustered doamins. The model validation process was conducted via benchmarking its predictions against the obtained data from the literature, which showed acceptable agreement and proved the model to be a quite accurate interpretation tool.