Abstract

Abstract

Abstract

Abstract
Tampered multimedia content is being increasingly used in a broad range of cybercrime activities. The spread of fake news, misinformation, digital kidnapping, and ransomware-related crimes are amongst the most recurrent crimes in which manipulated digital photos and videos are the perpetrating and disseminating medium. Criminal investigation has been challenged in applying machine learning techniques to automatically distinguish between fake and genuine seized photos and videos. Despite the pertinent need for manual validation, easy-to-use platforms for digital forensics are essential to automate and facilitate the detection of tampered content and to help criminal investigators with their work. This paper presents a machine learning Support Vector Machines (SVM) based method to distinguish between genuine and fake multimedia files, namely digital photos and videos, which may indicate the presence of deepfake content. The method was implemented in Python and integrated as new modules in the widely used digital forensics application Autopsy. The implemented approach extracts a set of simple features resulting from the application of a Discrete Fourier Transform (DFT) to digital photos and video frames. The model was evaluated with a large dataset of classified multimedia files containing both legitimate and fake photos and frames extracted from videos. Regarding deepfake detection in videos, the Celeb-DFv1 dataset was used, featuring 590 original videos collected from YouTube, and covering different subjects. The results obtained with the 5-fold cross-validation outperformed those SVM-based methods documented in the literature, by achieving an average F1-score of 99.53%, 79.55%, and 89.10%, respectively for photos, videos, and a mixture of both types of content. A benchmark with state-of-the-art methods was also done, by comparing the proposed SVM method with deep learning approaches, namely Convolutional Neural Networks (CNN). Despite CNN having outperformed the proposed DFT-SVM compound method, the competitiveness of the results attained by DFT-SVM and the substantially reduced processing time make it appropriate to be implemented and embedded into Autopsy modules, by predicting the level of fakeness calculated for each analyzed multimedia file.

Abstract
Deepfake and manipulated digital photos and videos are being increasingly used in a myriad of cybercrimes. Ransomware, the dissemination of fake news, and digital kidnapping-related crimes are the most recurrent, in which tampered multimedia content has been the primordial disseminating vehicle. Digital forensic analysis tools are being widely used by criminal investigations to automate the identification of digital evidence in seized electronic equipment. The number of files to be processed and the complexity of the crimes under analysis have highlighted the need to employ efficient digital forensics techniques grounded on state-of-the-art technologies. Machine Learning (ML) researchers have been challenged to apply techniques and methods to improve the automatic detection of manipulated multimedia content. However, the implementation of such methods have not yet been massively incorporated into digital forensic tools, mostly due to the lack of realistic and well-structured datasets of photos and videos. The diversity and richness of the datasets are crucial to benchmark the ML models and to evaluate their appropriateness to be applied in real-world digital forensics applications. An example is the development of third-party modules for the widely used Autopsy digital forensic application. This paper presents a dataset obtained by extracting a set of simple features from genuine and manipulated photos and videos, which are part of state-of-the-art existing datasets. The resulting dataset is balanced, and each entry comprises a label and a vector of numeric values corresponding to the features extracted through a Discrete Fourier Transform (DFT). The dataset is available in a GitHub repository, and the total amount of photos and video frames is 40,588 and 12,400, respectively. The dataset was validated and benchmarked with deep learning Convolutional Neural Networks (CNN) and Support Vector Machines (SVM) methods; however, a plethora of other existing ones can be applied. Generically, the results show a better F1-score for CNN when comparing with SVM, both for photos and videos processing. CNN achieved an F1-score of 0.9968 and 0.8415 for photos and videos, respectively. Regarding SVM, the results obtained with 5-fold cross-validation are 0.9953 and 0.7955, respectively, for photos and videos processing. A set of methods written in Python is available for the researchers, namely to preprocess and extract the features from the original photos and videos files and to build the training and testing sets. Additional methods are also available to convert the original PKL files into CSV and TXT, which gives more flexibility for the ML researchers to use the dataset on existing ML frameworks and tools.