Hybrid CMNV2: DeepFake faces classification and recognition using deep learning methods

• Introduced new hybrid deep learning model (CMNV2) for effective and efficient deepfake detection for machine learning applications. • Proposed a hybrid CMNV2 model for deepfake detection. • Evaluated 14 deep learning models on 5,000 images. • Achieved 99.10% accuracy, outperforming 13 models. • Demonstrated how the proposed model is applicable to real-world applications such as intelligent attendance and healthcare. Advanced image editing software makes it easy to generate fake images, such as combining faces. Improved binary categorization of real and deepfake faces is a major problem. We tested 14 facial detection methods: MobileNet, MobileNetV2, NASNetMobile, ResNet50V2, DenseNet121, ResNet101V2, DenseNet169, DenseNet201, InceptionV3, ResNet152V2, Xception, InceptionResNetV2, EfficientNetV2M and our model. We developed a hybrid “CMNV2” model using an improved MobileNetV2+convolutional architecture for fast feature embedding (CAFFE) after studying several deepfake classification model methods. This research added 5 layers to pre-trained model structures to improve CMNV2 model detection and classification of real and deepfake faces in photo images. This study uses Python models to train and test binary classification of 5,000 images into 2 categories: real and deepfake faces. Our CMNV2 Proposed model, which uses a combination of deep neural networks (DNNs), transfer learning (TL), and deep learning (DL) architecture, extracted image features for face detection (FD) and face classification (FC) better than 13 baseline CNN architectures approaches. Our CMNV2 model outperformed other 13 baseline CNN architectures deepfake classification models with 9 performance metrics evaluation, achieving 99.10% accuracy, 0.90% error rate, with better f1-score, computational time, precision, total parameters, Frozen (not updated during training), trained, and recall.