Deepfakes: implicações, detecção e prevenção

Abstract

With the rapid evolution of both hardware and software technologies, the creation of deepfake videos, synthetic videos manipulating people's images to cause harm—has become increasingly widespread; what began on social media for humorous purposes has now shifted toward criminal use. This study aims to analyze, test, and fine-tune Deep Neural Network-based models to identify deepfakes. Social media served as the primary source for collecting the real and manipulated videos used to build the ED-DeepForensics dataset (comprising 200 real and 240 fake videos). This dataset was used to train and test a hybrid detection architecture, in which pre-trained models underwent fine-tuning, a process that leverages existing training while incorporating new data to enhance performance. The architecture combines a ResNext Convolutional Neural Network (to analyze spatial/image data) with an LSTM Recurrent Neural Network (to analyze temporal/sequence data). Three analysis stages were successfully conducted using different frame sequences (10, 20, 40, and 80); the 80-frame sequence yielded the best accuracy (78.82%). The results demonstrate the classifier's convergence and improvement through training with new videos; however, future work should evaluate the model's limitations, even when trained on larger, more diverse datasets, and explore combining it with other models to achieve higher classification quality.