Deepfake is an environment in which AI technology is used to manipulate original digital videos, making them appear real. This poses a serious issue regarding the trustworthiness of digital data. The goal of the study is to create a reliable method to detect deepfakes using enhanced learning models named EfficientB0 and RESNET50. Frames are extracted from videos, and a HAAR cascade is applied to locate the face region, which is then sent as a dataset to train the model. This study utilized an open dataset from Kaggle to perform the experiment. The performance of the study is quantitatively measured using F1-score, accuracy, recall, and precision. The experiment showed that the hybrid model achieved superior prediction results of 89.08% compared to the standalone models. Hence, this study confirms that the proposed model works well to identify fake videos, which may help increase trust in digital data.
Image classification has evolved greatly in recent years, owing to the development of machine learning algorithms and the availability of large-scale image datasets. Convolutional Neural Networks (CNNs) have profoundly impacted the field of image classification due to their ability to learn hierarchical representations directly from pixel data. Unlike traditional machine learning algorithms, which rely on handcrafted features, CNNs can extract information hierarchically through multiple convolutional and pooling layers. Initially, compared various feature extraction techniques such as Histogram of Oriented Gradients (HOG), Local Binary Pattern (LBP), Scale Invariant-Future Transform (SIFT), Speeded-Up Robust Features (SURF), and wavelet domains like Discrete Cosine Transform (DCT) and Discrete Wavelet Transform (DWT) with standard classifiers. In order to improve the results, design a novel CNN model for image classification tasks, with an emphasis on hyperparameter optimization, data augmentation, dropout regularization, and efficient data loading. Experimental results on benchmark datasets show that our proposed CNN model outperforms baseline models, with an improving accuracy of 76.00%. These findings demonstrate the usefulness of our approach in advancing the state-of-the-art forward in image classification task.