A study ^[1][62] proposed LipForensics, a detection method for detecting forged face videos. LipForensics focuses on high-level semantic anomalies in mouth motions, which are prevalent in many created films. For training, the study used the FaceForensics++ (FF++) dataset, which includes 1.8 million modified frames and 4000 false videos created by two face-swapping algorithms, DeepFakes (DF) and FaceSwap (FS), as well as two face reenactment methods, Face2Face and NeuralTextures (NT). DeeperForensics (DFo) and FaceShifter (FSh), face-swapping, Celeb-DF-v2 (CDF), and DeepFake Detection Challenge (DFDC) datasets are used as a testing dataset. The study obtained 82.4%, 73.5%, 97.1%, and 97.6% accuracy from CDF, DFDC, FSh, and DFo, respectively.

A study ^[2][63] proposed deepfake detection as a fine-grained classification issue and presented a novel multi-attentional deepfake detection network. The study is composed of three major components. First, numerous spatial attention heads to direct the network’s attention to distinct local portions, second, textural feature enhancement block to zoom in on subtle artifacts in shallow features, and, third, combining low-level textural and high-level semantic characteristics driven by attention maps. The study used FF++, DFDC, and CDF datasets, where FF++ achieved 97.60% accuracy, CDF achieved 67.44% accuracy, and DFDC achieved 0.1679 Logloss.

Along the same lines, ^[3][64] proposes Multi-Feature Fusion Network (MFF-Net), a deepfake detection system that combines RGB features, and textural information extracted by an NN and signal processing techniques. The proposed system is composed of four major modules: (1) a feature extraction module to extract textural and frequency information, (2) a texture augmentation module to zoom the subtle textural features in shallow layers, (3) an attention module, and (4) two instances of feature fusion. Feature fusion includes fusing textural features from the shallow RGB branch and feature extraction module and fusing the textural features and semantic information. For the experimental process, the study used DFD, CDF, and FF++ datasets where FF++ achieved 99.73% accuracy, DFD achieved 92.53% accuracy, and CDF achieved 75.07% accuracy.

The authors propose Fake-Buster in ^[4][65], to address the issue of detecting face modification in video sequences using recent facial manipulation techniques. The study used NN compression techniques such as pruning and knowledge distillation to construct a lightweight system capable of swiftly processing video streams. The proposed technique employs two networks: a face recognition network and a manipulation recognition network. The study used the DFDC dataset which contains 119,154 training videos, 4000 validation videos, and 5000 testing videos. The study achieved 93.9% accuracy. Another study ^[5][66] proposed media forensics for deepfake detection using hand-crafted features. To build deepfake detection, the study explores three sets of hand-crafted features and three distinct fusion algorithms. These features examine the blinking behavior, the texture of the mouth region, and the degree of texture in the picture foreground. The study used TIMIT-DF, DFD, and CDF datasets. The evaluation results are obtained by using five fusion operators: concatenation of all features (feature-level fusion), simple majority voting (decision-level fusion), decision-level fusion weighted based on accuracy using TIMIT-DF for training, decision-level fusion weighted based on accuracy using DFD for training, and decision-level fusion weighted based on accuracy using CDF for training. The study concludes that hand-crafted features achieved 96% accuracy.

A lightweight 3D CNN is proposed in ^[6][67]. This framework involves the use of 3D CNNs for their outstanding learning capacity in integrating spatial features in the time dimension and employs a channel transformation (CT) module to reduce the number of parameters while learning deeper levels of the extracted features. To boost the speed of the spatial–temporal module, spatial rich model (SRM) features are also used to disclose the textures of the frames. For experiments, the study used FF++, DeepFake-TIMIT, DeepFake Detection Challenge Preview (DFDC-pre), and CDF datasets. The study achieved 99.83%, 99.28%, 99.60%, 93.98%, and 98.07% accuracy scores using FF++, TIMIT HQ, TIMIT LQ, DFDC-pre, and CDF datasets, respectively.

Several approaches have been presented to detect deepfake videos based on facial landmarks, e.g., Dlib ^[7][68] detector and multi-task convolutional neural network (CNN) ^[8][9][69,70], etc. For example, a deep learning model has been proposed by ^[10][71] to detect deepfake videos. The study uses the FaceForensics dataset to train a CNN. The FaceForensics++ dataset comprises real and fake videos, of which 363 videos are real videos and 3068 are fake videos. Additionally, the model is tested with a variety of AI approaches including layer-wise relevance propagation (LRP) and local interpretable model-agnostic explanations (LIMEs) to provide clear representations of the prominent parts of the picture as pointed out by the model. Firstly, the face is extracted from the images using Dlib. After that, CNN XceptionNet is used to extract the features. To analyze the results of the models, the study utilizes the Xception network which is a conventional CNN with depth-wise separable convolutions (DWSCs) and LIME. Using 1.3× and 2× background scales, the study trains the algorithm with 90.17% accuracy on the image dataset. With the experimental process at 1.3×, the model achieves 94.33% accuracy, and at 2× the model achieves 90.17% accuracy.

A study ^[11][72] proposes a system for detecting deepfake videos using a support vector machine (SVM). For detecting fake videos, the model utilizes feature points (FPs) taken from the video to build an AI model. Different FP extraction methods have been used for experiments including histogram of oriented gradient (HOG), oriented features from accelerated segment test (FAST), rotated binary robust independent elementary features (BRIEF), oriented robust binary (ORB), binary robust invariant scalable key-points (BRISK), KAZE, speeded-up robust features (SURF). The study uses the dataset from ^[12][73] which comprises 90 MP4 videos with a length of about 30 s. Half of the videos in the collection are fake while the other half are real. The HOG FP extraction method obtains a 95% accuracy whereas ORB achieves 91% accuracy, SURF achieves 90.5% accuracy, BRISK 87%, FAST 86.5%, and KAZE 76.5% accuracy. The results show that the HOG extracted FP shows the highest accuracy.

Xin et al. ^[13][74] propose a deepfake detection system based on inconsistencies found in head poses. The study focuses on how deepfakes are created by splicing a synthetic face area into the original picture, as well as how it can employ 3D posture estimation for identifying manufactured movies. The study points out that the algorithms are employed to generate various people’s faces but without altering their original expressions. This leads to mismatched facial landmarks and facial features. As a result of this deepfake method, the landmark positions of a few fake faces may sometimes differ from those of the actual ones. They may be distinguished from one other based on the difference in the distribution of the cosine distance between their two head orientation vectors. The study uses Dlib to recognize faces and retrieve 68 facial landmarks. OpenFace2 is used to build a standard face 3D model, from which a difference calculation is made based on that model. The suggested system makes use of the UADFV data. In this case, an SVM classifier using radian basis function (RBF) kernels is utilized. When using an SVM to classify the UADFV dataset, the SVM classifier achieved an area under the receiver operating characteristic curve (AUROC) of 0.89.

An automated deepfake video detection pipeline based on temporal awareness is proposed in ^[14][75]. For this purpose, the study proposes a two-stage analysis where the first stage involves using a CNN to extract characteristics at the frame level, followed by a recurrent neural network (RNN) that can detect temporal irregularities produced during the face-swapping procedure. The study created a dataset of 600 videos, half of which are gathered from a variety of video-hosting websites, while the other 300 are random selections from the HOHA dataset. With the use of sub-sequences of n=20,40,80 frames, the performance of the proposed model is evaluated in terms of detection accuracy. The results show that from the selected frames, CNN-LSTM achieves the highest accuracy of 97.1% from 40 and 80 frames.

Several patterns and clues may be used to investigate spatial and temporal information in deepfake videos. For example, a study ^[15][76] created FSSPOTTER to detect swapped faces in videos. The spatial feature extractor (SFE) is used to distribute the videos into multiple segments, each of which comprises a specified number of frames. Input clips are sent into the SFE, which creates frame-level features based on the clips’ frames. Visual geometry group (VGG16) convolution layers with batch normalization are used as the backbone network to extract spatial information in the intra-frames of the image. The superpixel-wise binary classification unit (SPBCU) is also used with the backbone network to retrieve additional features. An LSTM is used by the temporal feature aggregator (TFG) to detect temporal anomalies inside the frame. In the study, the probabilities are computed by using a fully connected layer, as well as a softmax layer to determine if the clip is real or false. For the evaluation process, the study uses FaceForensics++, Deepfake TIMIT, UADVF, and Celeb-DF datasets. FSSPOTTER achieves a 91.1% accuracy from UADFV, Celeb-DF 77.6%, Deepfake TIMITHQ 98.5%, and LQ 99.5%, whereas using FaceForensics++, FSSPOTTER obtains 100% accuracy. In ^[16][77], the CNN-LSTM combo is used to identify and classify the videos into fake and real. Deepfake detection (DFD), Celeb-DF, and deepfake detection challenge (DFDC) are the datasets utilized in the analysis. The experimental process is performed with and without the usage of transfer learning. The XceptionNet CNN is used for detection. The study combined all three datasets to make predictions. Using the proposed model on the combined dataset, an accuracy of 79.62% is achieved without transfer learning, and with transfer learning the accuracy is 86.49%.

A YOLO-CNN-XGBoost model is presented in ^[17][10] for deepfake detection. It incorporates a CNN, extreme gradient boosting (XGB), and the face detector you only look once (YOLO). As the YOLO face detector extracts faces from video frames, the study uses the InceptionResNetV2 CNN to extract facial features from the extracted faces. The study uses the CelebDF-FaceForencics++ (c23) dataset which is a combination of two popular datasets: Celeb-DF and FaceForencics++ (c23). Accuracy, specificity, precision, recall, sensitivity, and F1 score are used as evaluation parameters. Results indicate that the CelebDF-FaceForencics++ (c23) combined dataset achieves a 90.62% area under the curve (AUC), 90.73% accuracy, 93.53% specificity, 85.39% sensitivity, 85.39% recall, and 87.36% precision. The model obtains an average F1 score of 86.36% for the combined dataset.

Besides using the image processing approaches for detecting deepfake videos, physiological signals can also be used for the same purpose. For example, a study ^[18][78] proposed a system to identify deepfake videos using mouth movements. For this purpose, the study used a CNN and designed a deepfake detection model with mouth features (DFT-MF). To detect deepfakes, two different datasets are used, the deepfake forensics dataset and the VID-TIMIT dataset, that contain real and fake videos. The deepfake forensic dataset comprises a total of 1203 videos with 408 real videos and 795 fake videos, whereas the VID-TIMID dataset comprises 320 low-quality (LQ) and 320 high-quality (HQ) videos. In the preprocessing step, a Dlib classifier is utilized to recognize face landmarks. As an example, the face according to Dlib has (49, 68) coordinates. A person’s eyebrows, nose, and other facial features can be determined using the Dlib library. Afterward, all frames with a closed mouth are excluded by measuring the space between lips. According to the suggested model, sentence length is determined by the number of words that are spoken. The study also focuses on the rate of speech and shows that about 120–150 words per minute are spoken. In the proposed system, deepfake videos are determined by facial emotions and speech speed. The experimental results of DFT-MF show that, using the deepfake forensics dataset, a 71.25% accuracy can be obtained, whereas with the deepfake Vid-TIMIT dataset, DFT-MF achieves a 98.7% accuracy from LQ and 73.1% accuracy from HQ.

A study ^[19][79] leverages a deep neural network (DNN) model to study fake videos and formulate a novel approach for exposing fraudulent face videos. An eye blinking signal is detected in the videos, which is a physiological signal that does not show up well in synthetically created false videos. CNN and eye aspect ratio methods are employed in the long-term recurrent convolutional neural network (LRCN) model. For this purpose, an eye blinking video dataset of 50 videos of a 30-second duration was generated. The study utilized 40 movies for training the LRCN model and 10 for testing. Results show that regarding area under reciprocal control (AUC), the result shows that LRCN performs best with 99% accuracy whereas 98% and 79% accuracy is achieved by the CNN and eye aspect ratio (EAR), respectively.

‘DeepVision’, a novel algorithm to discriminate between real and fake videos, is presented in ^[20][80] and utilizes eye blink patterns for deepfake video detection. Fast-hyperFace and EAR are used to recognize the face and calculate the eye aspect ratio. The study created a dataset based on eye blinking patterns for experiments. The features of eye blink count and eye blink duration are retrieved to determine if a video is real or a deepfake. Experimental results using DeepVision show that an accuracy of 87.5% can be obtained.

Korshunov and Marcel ^[21][1] studied baseline methods based on the discrepancies between mouth movements and voice, as well as many versions of image-based systems frequently employed in biometrics to identify deepfakes. The study found that auditory and visible features can be used for mouth movement profiles in deepfakes. An RNN based on LSTM is utilized to recognize real and fake videos and principal component analysis (PCA) and latent Dirichlet allocation (LDA) have been utilized to decrease the dimensions of the blocks of data. In the second case, i.e., voice movements, the study used two detection methods, raw faces as features and image quality measurements (IQMs). For this purpose, 129 features were investigated including signal-to-noise ratio, specularity, blurriness, etc. The final categorization was based on PCA-LDA or SVM. For the deepfake TIMIT database, the study suggested that the detection techniques based on IQM+SVM produced the best results of 3.3% low-quality energy efficiency ratio (LQ EER) and 8.9% high-quality EER (HQ EER).

Biological signals have been predominantly used in the medical field to determine the physical and emotional state of people ^[22][81]. Using the features from the data indicating heart rate, galvanic skin response, electrocardiogram, etc., abnormal biological signals can be identified by experts. For medical analysis, such approaches require the use of sensors and nodes which are placed on different limbs of the human body; this is not possible for deepfake detection. Intuitively, computer experts have designed algorithms that can measure biological signals using features from the video data such as changes in color, motion, subtle head movements, etc. ^[23][82].

Besides the physiological signals gathered from the videos, biological signals present a potential opportunity to identify deepfakes. For example, a study ^[24][83] detected deepfakes from videos using a ‘FakeCatcher’ system. The study proposed a FakeCatcher technique for detecting synthesized information of portrait videos as a deep fake prevention solution. This method is based on the findings that biological signals collected from face regions are poorly retained geographically and temporally in synthetic content. Different methods are proposed for enhancements to the derived PPG signal’s quality, as well as the extraction process’s reliability. Chrominance attributes ^[25][84], green channel elements ^[26][85], optical properties ^[27][86], Kalman filters ^[28][87], and distinct face regions ^[26][27][29][85,86,88] are some of the suggested enhancements. The study used six biological signals, GL, GR, GM, CL, CR, and CM, where GL represents green left cheek, GR represents green right cheek, GM is the green mid-region, CL represents chrominance left, CR represents chrominance right, and CM represents chrominance mid-region. Experiments were performed using three benchmark datasets, FaceForensics, FaceForensics++, and CelebDF, in addition to the newly collected dataset, Deep Fakes (DF). The DF dataset comprises 142 portrait videos collected ’in the wild’ where each video has a length of 32 min. For detection, the study uses a CNN classifier trained on the above-mentioned features. Results indicate that the CNN achieves a 91.07% accuracy from the DF dataset, 96% accuracy for the FaceForensics dataset, 91.59% from the CelebDF dataset, and 94.65% using the FaceForensics++ dataset.

The authors of ^[30][89] discussed several biological signals for deepfake video detection, eye and gaze properties, by which deepfakes differ. Furthermore, the researchers combined those characteristics into signatures and compared original and fake videos, generating geometric, visual, metric, temporal, and spectral variances. The study used FaceForensics++, Deep Fakes, CelebDF, and DeeperForensics datasets. To categorize any video in the wild as false or real, the researchers used a deep neural network. With the proposed approach, 80.0% accuracy on FaceForensics++, 88.35% with Deep Fakes (in the wild), 99.27% using CelebDF, and 92.48% using the DeeperForensics dataset can be obtained.

The research described in ^[31][90] offers a method for not only distinguishing deepfakes from the original movies but also presents a generative model that underpins a deepfake. DL techniques are used to categorize deepfakes using CNNs. The authors found that such manipulative artifacts from biological signals can be used to detect deepfakes. The findings reveal that spatial–temporal patterns in biological signals may be thought of as a representative projection of residuals. The results show that the method correctly detects bogus films with a 97.29% accuracy and correctly detects the source model with 93.39% accuracy.

Unlike the detection of video deepfakes, which are sequences of images, deepfake image detection focuses on identifying a single image as a deepfake. For example, the study described in ^[32][91] proposed a system to detect deepfake human faces. The expectation-maximization (EM) method was used to extract features from the image. For the classification, k-nearest neighbors (k-NN), SVM, and LDA algorithms were applied. In the study, deepfake images were generated using the based approach. Fake pictures were created by five different GAN techniques, AttGAN, StarGAN, GDWCT, StyleGAN, and StyleGAN2, with the CelebA dataset as ground truth for non-fakes. For experiments, 6005 images from AttGAN, 3369 images from GDWCT, 9999 images from StyleGAN, 5648 images from StarGAN, and 3000 images from StyleGAN2 were used. The study achieved the best accuracy of 99.81% on StyleGAN2-generated images with linear SVM.

A comprehensive evaluation of face manipulation techniques was conducted in ^[33][93] using a variety of modern detection technologies and experimental settings, including both controlled and uncontrolled scenarios. The study used four distinct deepfake image databases using different GAN variants. The StyleGAN architecture was used to create 150,000 fake faces collected online. Similarly, the 100K-faces public database that contains 80,000 synthetic faces was used. The GANprintR approach was used to remove GAN fingerprint information from the iFakeFaceDB database, which is an enhanced version of previous fake databases. Findings of the study reveal that an EER of 0.02% is obtained in controlled situations which is similar to the best recent research. From the iFakeFaceDB dataset, the study achieved 4.5% EER for the best fake detectors.

A method for detecting fake images was developed by Dang et al. ^[34][94]. An attention process was used to enhance the functionality of feature maps for the CNN model. Fake pictures were produced using the FaceApp software, which has up to 28 distinct filters including age, color, glasses, hair, etc. Similarly, the StarGAN method was used which has up to 40 different filters ^[35][95]. The CNN model was also tested on the study’s own collected DFFD dataset with 18,416 real and 79,960 fake pictures produced using FaceApp and StarGAN. The results were outstanding, with an EER of less than 1.0% and 99.9 percent AUC.

Along the same direction, Wang et al. ^[36][96] used the publicly accessible commercial software Face-Aware Liquify tool from Adobe Photoshop to create new faces. Additionally, skilled artists used 50 real images to produce modified images. Participants were shown fake and real photos and asked to categorize the images into groups, as part of Amazon Mechanical Turk (AMT) research. Humans were able to attain only 53.5% accuracy, which is close to chance (50 percent). Two alternative automated methods were presented against the human study: One using dilated residual networks (DRNs) to estimate whether or not the face has been distorted, and another using the optical flow field to detect where manipulation has occurred and reverse it. Using the automatic face synthesis manipulation, the study achieved 99.8% accuracy, and using the manual face synthesis manipulation, the study achieved 97.4% accuracy.

A study ^[37][97] used CNN models to detect fake face images. Different CNN techniques were used for this purpose, such as VGG16, VGG19, ResNet, and XceptionNet. For this purpose, the study used two datasets for manipulated and original images. For real images, the study used the CelebA database whereas for the fake pictures two different options were utilized. First, machine learning techniques based on GAN were used in ProGAN, and secondly, manual approaches were leveraged using Adobe Photoshop CS6 based on different features such as cosmetics, glasses, sunglasses, hair, and headwear alterations, among other things. For experiments, a range of picture sizes (from 32 × 32 to 256 × 256 pixels) were tested. A 99.99% accuracy was achieved within the machine-created scenario, whereas 74.9% accuracy was achieved from the CNN model. Another study ^[38][98] suggested detection methods for fake images using visual features such as eye color and missing details for eyes, dental regions, and reflections. Machine learning algorithms, logistic regression (LR) model and multi-layer perceptron (MLP), were used to detect the fake faces. The proposed technique was evaluated on a private FaceForensics database where LR achieved an 86.6% accuracy and the MLP achieved an 82.3% accuracy.

A restricted Boltzmann machine (RBM) is used in ^[39][99] to develop deepfake images made using facial image digital retouching. By learning discriminative characteristics, each image was classified as original or fake. The authors generated two datasets for fake images using the actual ND-IIITD Retouching (ND-IIITDR) dataset (collection B) and Celebrity Retouching (CR) which is a set of celebrity facial pictures retrieved from the internet. Fake pictures were created with the help of Max25’s PortraitPro Studio software, which took into account elements of the face such as the texture of the skin and skin tone, as well as eye coloration. In the CR and ND-IIITD Retouching datasets, the study achieved an accuracy of 96.2% and 87.1% percent, respectively.

A study ^[40][100] proposed a face X-ray, a novel image representation for identifying fraudulent face images or deepfakes. The basic finding for face X-rays is that most current face alteration algorithms share a similar step of blending a changed face into an existing backdrop picture, and there are inherent image disparities across the blending boundaries. The study used FF++ and Blending Images (BI) datasets for training and DFD, DFDC, and CDF datasets for testing. For the experimental process, the study used a CNN and achieved 95.40% accuracy from DFD, 80.92% accuracy from DFDC, and 80.58% accuracy from CDF.

The Fake or Real (FoR) dataset ^[41][101], which includes eight synthetically created English-accented voices using the Deep Voice 3 and Google-Wav Net generation models, was released in 2019. It is publicly available; its most important feature is that it contains sounds in two different formats, MP3 and WAV. The complete dataset consists of 198,000 files, comprising 111,000 original samples and 87,000 counterfeit samples, each lasting two seconds. The Arabic Diversified Audio ^[42][102] dataset (Ar-DAD), which was acquired via the Holy Quran audio site, was apparently a fake audio collection of Arabic speakers. The audio is of 30 male Arabian reciters and 12 mimics, and it comprises the original and mimicked voices of Quran reciters. The reciters are Arabians from Egypt, Sudan, Saudi Arabia, Yemen, Kuwait, and the United Arab Emirates. There are 379 false and 15,810 actual samples in the data, each voice has a 10 s duration.

The H-Voice dataset ^[43][103] was recently established using fake and real voices in various languages including French, English, Portuguese, Spanish, and Tagalog. It includes samples stored in the PNG format as a histogram. There are 6672 samples in this dataset and it is organized into six folders: ‘training original’, ‘training fake’, ‘validation original’, ‘validation fake’, ‘external test 1’, and ‘external test 2’. Each category has a different number of samples, where the first category has 2020 histograms while the second category contains 2088 histograms, 2016 imitations, and 72 deep voices. The third category contains 864 histograms, ‘validation fake’ contains 864 histograms, and ‘external test 1’ and ‘external test 2’ are further divided into two sub-folders, ‘fake’ and ‘original’. The ‘external test 1’ set contains a total of 760 histograms (380 fake imitation histograms and 380 original histograms) while the ‘external test 2’ set contains 76 histograms (72 fake deep voice histograms and four original histograms).

Finally, the ASV spoof 2021 challenge dataset includes two false circumstances, one cognitive and one actual. False audio is created in the cognitive environment utilizing synthetic software, whereas fake audio is created in the actual environment by replicating prerecorded sounds using sections of genuine speaker data. This dataset has not yet been released; prior versions are freely available (2015 ^[44][104], 2017 ^[45][105], and 2019 ^[46][106]).

A large variety of methods and techniques for creating fake audio have prompted a wide interest in detecting deepfake audio in many languages. This section presents the works on recognizing mimicked and synthetically created voices. In general, there are two types of techniques that are used currently: ML and DL approaches.

Traditional ML methods are commonly used in the identification of fake audios. A study ^[47][107] created a own fake audio dataset by extracting entropy features using an imitation technique named the H-Voice dataset ^[43][103]. To distinguish between the fake and real audio, the study used ML model LR. LR achieved 98% accuracy for real vs. fake audio detection. The study points out that manual feature extraction can boost the performance of the proposed approach.

To identify artificial audio from natural human voices, Singh et al. ^[48][108] used the H-Voice dataset and suggested a quadratic SVM (QSVM) method. The study classified the audio into two types, human and AI-generated. Additional ML approaches including linear discriminant (LD), quadratic LDSVM, weighted KNN, boosted tree ensemble, and LR were compared against this model. It is observed that QSVM beats other traditional approaches by 97.56% accuracy and has only a 2.43% misclassification rate. Similarly, Borrelli et al. ^[49][109] created an SVM model using RF to classify artificial voices using a novel audio component known as short-term long-term (STLT). The Automatic Speaker Verification (ASV) spoof 2019 challenge dataset was used to train the models. The results show that RF performs best compared to SVM with a 71% accuracy result. In a similar way, ^[50][110] also used the H-Voice dataset and compared the effectiveness of SVM with the DL technique CNN to distinguish fake audio from actual stereo audio. The study discovered that the CNN is more resilient than the SVM, even though both obtained a high classification accuracy of 99%. The SVM, however, suffers from the same feature extraction issues as the LR model did.

A study ^[51][111] designed a CNN method in which the audio was converted to scatter plot pictures of surrounding samples before being input into the CNN model. The generated method was evaluated using the Fake or Real (FoR) dataset and achieved a prediction accuracy of 88.9%. Whereas the suggested model solved the generalization problem of DL-based architectures by training with various data generation techniques, it did not perform well as compared to other models in the literature. The accuracy and equal error rate (EER) were 88% and 11%, respectively, which are lower than other DL models.

Another similar study is ^[52][112] that presented deep sonar and a DNN model. The study presented the neuron behaviors of speaker recognition (SR) systems in the face of AI-produced fake sounds. In the classification challenge, their model is based on layer-wise neuron activities. For this purpose, the study used the voices of English speakers from the FoR dataset. Experimental results show that 98.1% accuracy can be achieved using the proposed approach. The efficiency of the CNN and BiLSTM was compared with ML models in ^[53][113]. The proposed approach detects imitation-based fakeness from the Ar-DAD of Quranic audio samples. The study tested the CNN and BiLSTM to identify fake and real voices. SVM, SVM-linear, radial basis function (SVMRBF), LR, DT, RF, and XGBoost were also investigated as ML algorithms. The research concludes that the SVM has a maximum accuracy of 99%, while DT has the lowest accuracy of 73.33%. Furthermore, the CNN achieves a 94.33% detection rate which is higher than BiLSTM.

Similar to deepfake videos and images that are posted online as separate units, tweets posted on Twitter may also be fake, so they are also called deepfakes. Therefore, a specialized study ^[54][114] focused on detecting deepfakes from tweets alone. The study collected a dataset of deepfake tweets named the TweepFake dataset. The study collected 25,572 randomly selected tweets from 17 human accounts imitated by 23 bots. Markov chains, RNN, RNN+Markov, and LSTM are some of the approaches used to create the bots. The study used 13 deepfake detection methods: LR_BOW, RR _BOW, SVC_BOW, LR_BERT, RF_BERT, SVC_BERT, CHAR_CNN, CHAR_GRU, CHAR_CNNGRU, BERT_FT, DISTILBERT_FT, ROBERTA_FT, and XLNET_FT. Experimental results show that ROBERTA_FT performs best with an 89.6% accuracy whereas LR_BOW achieved 80.4%, RF_BOW achieved 77.2%, SVC_BOW achieved 81.1%, LR_BERT achieved 83.5%, RF_BERT achieved 82.7%, SVC_BERT achieved 84.2%, CHAR_CNN achieved 85.1%, CHAR_GRU achieved 83%, CHAR_CNNGRU achieved 83.7%, BERT_FT achieved 89.1%, DISTILBERT_FT achieved 88.7%, and XLNET_FT achieved 87.7% accuracy.