Deepfake Face Detection Using LSTM and CNN

Abstract

The fast evolvement of deepfake introduction tech- nology is critically threating media facts trustworthiness. The consequences impacting focused individuals and establishments may be dire. In this work, we study the evolutions of deep studying architectures, especially CNNs and Transformers. We diagnosed 8 promising deep learning architectures, designed and evolved our deepfake detection fashions and conducted experiments over well-installed deepfake datasets. those datasets protected the modern 2nd and third generation deepfake datasets. We evaluated the effectiveness of our evolved unmarried version detectors in deepfake detection and move datasets evaluations. This have a look at introduces a comprehensive methodology for facial picture evaluation, starting with a cautiously curated dataset of photos in either ’.jpg’ or ’.png’ formats. This standard- ization is essential for the subsequent characteristic extraction technique, which makes a speciality of capturing important characteristics of the faces through local capabilities consisting of imply, widespread deviation, and variance. these statistical metrics provide a robust basis for information versions in facial attributes, enhancing the model’s potential to distinguish among diverse identities. In to addition improve the overall performance of the version, a Transformer architecture is leveraged for face detection, coupled with advanced data augmentation techniques that increase the dataset and help the model generalize better to unseen pix. The heart of the evaluation relies on a sophisticated deep getting to know framework that integrates Convolutional Neural Networks (CNN) and Long Short Term Memory(LSTM) classifiers. This hybrid method capitalizes at the strengths of each architectures: CNNs excel in spatial function extraction from pix, while LSTMs are adept at taking pictures temporal dependencies, making them appropriate for sequences of information. The dataset is judiciously cut up into schooling (90) and testing (10) subsets to facilitate effective model schooling and assessment. overall performance metrics, mainly accuracy and mistakes rates, are hired to assess the model’s effectiveness in face reputation responsibilities. by using analyzing these metrics, the take a look at provides valuable insights into the strengths and limitations of the proposed technique, laying the foundation for destiny upgrades in facial picture analysis and reputation technology.