With the exponential growth of high-quality fake images in social networks and media, it is necessary to develop recognition algorithms for this type of content. One of the most common types of image and video editing consists of duplicating areas of the image, known as the copy-move technique. Traditional image processing approaches manually look for patterns related to the duplicated content, limiting their use in mass data classification. In contrast, approaches based on deep learning have shown better performance and promising results, but they present generalization problems with a high dependence on training data and the need for appropriate selection of hyperparameters. To overcome this, we propose two approaches that use deep learning, a model by a custom architecture and a model by transfer learning. In each case, the impact of the depth of the network is analyzed in terms of precision (P), recall (R) and F1 score. Additionally, the problem of generalization is addressed with images from eight different open access datasets. Finally, the models are compared in terms of evaluation metrics, and training and inference times. The model by transfer learning of VGG-16 achieves metrics about 10% higher than the model by a custom architecture, however, it requires approximately twice as much inference time as the latter.

Recently, some state-of-the-art works have used deep learning-based architectures, specifically convolutional neural networks (CNNs), for banknote recognition and counterfeit detection with promising results. However, it is not clear which design strategy is more appropriate (custom or by transfer learning) in terms of classifier performance and inference times for massive data applications. This paper presents a comparison of the two design strategies in various types of architecture. For the transfer learning (TL) strategy, the most appropriate freezing points in CNN architectures (sequential, residual and Inception) are identified. In addition, a custom model based on an AlexNet-type sequential CNN is proposed. Both the TL and the custom models were trained and compared using a Colombian banknote dataset. According to the results, ResNet18 achieved the best accuracy, with 100%. On the other hand, the network with the shortest inference times was the proposed custom network, since its performance is up to 6.48-times faster in CPU and 16.29-times faster in GPU than the inference time with the models by transfer learning.

This paper presents H-Voice, a dataset of 6672 histograms of original and fake voice recordings obtained by the Imitation [1,2] and the Deep Voice [3] methods. The dataset is organized into six directories: Training_fake, Training_original, Validation_fake, Validation_original, External_test1, and External_test2. The training directories include 2088 histograms of fake voice recordings and 2020 histograms of original voice recordings. Each validation directory has 864 histograms obtained from fake voice recordings and original voice recordings. Finally, External_test1 has 760 histograms (380 from fake voice recordings obtained by the Imitation method and 380 from original voice recordings), and External_test2 has 76 histograms (72 from fake voice recordings obtained by the Deep Voice method and 4 from original voice recordings). With this dataset, the researchers can train, cross-validate and test classification models using machine learning techniques to identify fake voice recordings.