Deep learning (DL) has become one of the most influential fields in computer science, directly affecting today's human life and society. Like every other technological innovation in history, deep learning has been used for some illegal purposes. Deepfakes are such a deep learning application. Hundreds of studies have been conducted in the past few years to invent and optimize various Deepfake detection using AI. This article mainly discusses how to detect Deepfake.
To deal with deepfakes, deep learning methods as well as machine learning (non-deep learning) methods have been developed to detect them. Deep learning models need to consider a large number of parameters, so a large amount of data is required to train such models. This is exactly why DL methods have higher performance and accurate results compared to non-DL methods.
Most deepfake generators will leave some traces during the deepfake process. These changes in deepfake videos can be classified as spatial inconsistencies: incompatibilities that occur within individual frames of the video and temporal inconsistencies: incompatible features that appear in the sequence of video frames.
Spatial inconsistencies include areas of the face that are incompatible with the background of the video frame, resolution changes, and partially rendered organ and skin textures (which may not render all human features of the face correctly). Most common deepfake generators are unable to render features such as blinks and teeth. And sometimes white strips are used to replace the teeth that are even visible to the naked eye on the still frame (below).
Temporal inconsistencies include abnormal eye blinks, head postures, facial movements, and brightness changes in video frame sequences.
These traces, both spatial and temporal, left by deepfake generators can be identified by deepfake detectors made from deep neural networks (DNN). The widespread use of familiar generative adversarial networks (GANs) in deepfake generators challenges the balance between fake detection and generation.
Deepfake detector is a binary classification system that determines whether input digital media is real or fake. Deepfake detection is not performed by a single black-box-like module, but consists of several other modules and steps that work together to provide detection results. Common steps in Deepfake detection are as follows [2].
A typical DL-based deepfake detector contains 3 main components to perform the above tasks.
The main steps will be explained in detail next: data preprocessing, feature extraction and detection/classification process.
After the data collection phase, the data should be preprocessed before the training and testing steps for deepfake detection. Data preprocessing is done automatically using available libraries such as OpenCV python, MTCNN and YOLO etc.
Data augmentation also plays a crucial role in improving the performance of deepfake detectors. Augmentation techniques such as rescaling (stretching), shear mapping, scaling augmentation, rotation, brightness changes, and horizontal/vertical flipping can be applied to increase the generalization of the dataset [3].
The first step in data preprocessing is to extract individual frames from the video clip. After extracting the frames, you need to detect faces from the extracted video frames. Since facial regions often appear anomalies, selecting only facial regions helps the feature extraction model focus only on the region of interest (ROI), thus saving the computational cost for full-frame scanning. Once facial regions are detected, they are cropped from the rest of the background of the frame and follow a series of steps to make them available for model training and testing. Another reason to crop facial regions is to make all input images to the model the same size.
The frames preprocessed in the previous step will be sent to the feature extractor. Most feature extractors are based on convolutional neural networks (CNN). Some recent studies have demonstrated the effectiveness and efficiency improvement of the application of capsule networks in the feature extraction process, which is a new trend.
The feature extractor extracts the spatial features available on the preprocessed video frames. Feature extraction can extract visual features, local features/facial landmarks, such as the position of eyes, nose, mouth, dynamics of mouth shape, blinking and other biological features. The extracted feature vectors are then sent to the classifier network to output decisions.
The deep learning model used for classification is often called the backbone of the deepfake detector. As the name suggests, the classification network is responsible for the most important task in the deepfake detection pipeline: namely, classifying and determining the probability of whether the input video is a deepfake. Most classifiers are binary classifiers, where the deepfakes output is (0) and the original frame output is (1).
The classifier is yet another convolutional layer (CNN) or similar deep learning architecture such as LSTM or ViT. The actual capabilities of a classification model vary depending on the DNN used. For example, the blink features extracted in the feature extractor module can be used by the LSTM module in the classification module to determine the temporal inconsistency of the frame blink pattern and determine whether the input is a Deepfake [3]. In most cases, the last layer in a deepfake detector is a fully connected layer. Since the outputs of the convolutional layers represent high-level features of the data, these outputs are flattened and concatenated to a single output layer to produce the final decision.
Over the past few years, there have been significant developments in both the creation and detection of deepfakes. Research related to deepfake detection using deep learning technology has also made great progress due to the accuracy of the results compared to non-deep learning methods. Deep neural network architectures such as CNN, RNN, ViT, and capsule networks are widely used in the implementation of deepfake detectors. A common deepfake detection pipeline consists of a data preprocessing module, a CNN-based feature extractor, and a classification module.
In addition, deepfake detection has a strong dependence on the traces left by the deepfake generator on the deepfake. Since current GAN-based deepfake generators are capable of synthesizing more realistic deepfakes with minimal inconsistencies, new methods must be developed to optimize deepfake detection. Deepfake detection methods based on deep ensemble learning techniques can be considered as modern and comprehensive methods to combat deepfakes [4]. Nonetheless, a gap for effective and efficient deepfake detectors still exists.
The above is the detailed content of Overview of Deepfake Detection Based on Deep Learning. For more information, please follow other related articles on the PHP Chinese website!
![[Web front-end] Node.js quick start](https://img.php.cn/upload/course/000/000/067/662b5d34ba7c0227.png)
