Deepfake forensics: a survey of digital forensic methods for multimodal deepfake identification on social media

Prevalent data sets

We begin by discussing the datasets currently most commonly applied for detecting deepfakes using methodologies rooted in deep learning. We explore various and latest datasets extensively used by models designed to detect deepfakes.

The CelebA dataset

The dataset discussed in Liu et al. (2018) is extensive, comprising 200,000 photographs of well-known individuals. Each image in the collection contains 40 distinct attributes that provide labels for various characteristics. The dataset has many photos that exhibit various stances and backdrop disturbances. As a result, it serves as a valuable resource due to its extensive diversity, large quantity, and comprehensive annotations. These annotations include 10,177 unique identities, 202,599 facial images, five landmark positions, and 40 binary attribute annotations for each image.

DeeperForensics-1.0 dataset

The DeeperForensics-1.0 dataset (Liming et al., 2020) is an extensive, heterogeneous, and exemplary repository that facilitates the detection of falsifications. The database contains a collection of 60,000 films and 17.6 million frames with automated face swaps, all captured at a resolution of 1920x1080 pixels. The first film was sourced from a diverse group of 100 performers from 26 nations, exhibiting a wide range of skin tones and ages 20 to 45. The dataset includes eight distinct natural emotions (anger, fear, joy, disgust, surprise, contempt, sadness, and neutrality) taken from various perspectives, ranging from −90 to +90 degrees. The unusual stances, expressions, and lighting of the source photographs highly influence the quality of the collection.

Deepfake-TIMIT dataset

The Deepfake-TIMIT dataset (Korshunov & Sebastien, 2018) is categorized into videos of poor quality and high quality. The segment designated as low-quality comprises a total of 320 movies, with an approximate frame count of 200 frames per video. Each frame has a resolution of 64 × 64 pixels. On the other hand, the high-quality section consists of 320 image sequences, with an average frame count of 400 frames per sequence. The resolution of each frame in this section is 128 × 128 pixels.

Celeb-DF dataset

The Celeb-DF dataset (Li et al., 2020) is an extensive compilation of modified videos that prominently showcase celebrities sourced from their YouTube output. The dataset consists of 5,639 films of good quality, with a total of almost two million frames. Each frame has a size of 256 × 256 pixels. The collection encompasses various celebrities representing different nationalities and age groups, encompassing both male and female individuals. Each movie lasts around 13 s, with a 30 frames per second frame rate. These videos encompass a range of situations, such as different orientations, face sizes, lighting settings, and backgrounds.

FaceForensics dataset

The dataset known as FaceForensics (FF) (Andreas et al., 2018) contains around 500,000 modified images obtained from a collection of 1,004 videos. The subject matter is separated into two groups, which have been established using the Face2Face reenactment methodology. The datasets encompass a Source-to-Target Reenactment Dataset, which facilitates reenactments between two arbitrarily chosen movies, and a Self Reenactment Dataset, which employs a single video as both the source and the target. The dataset consists of 1,408 training films, 300 validation videos, and 300 test videos, containing 732,391,151,835 and 156,307 pictures, respectively.

UADFV dataset

The UADFV dataset by Li, Ming-Ching & Siwei (2018) is a synthetic resource developed by the University of Albany. Its purpose is to aid in identifying deepfake movies by utilizing physiological cues, particularly the analysis of blinking patterns. The dataset consists of 49 fabricated films created using the FakeApp software. Each video has a resolution of 294 by 500 pixels and an average duration of 11.14 s.

Faceforensics++ dataset

The FaceForensics++ (FF++) dataset by Rossler et al. (2019) builds upon the FaceForensics dataset, offering a publicly available benchmark for detecting realistic artificial facial images. The dataset consists of 1,000 carefully curated videos, predominantly from YouTube, featuring roughly 60% male and 40% female subjects. Regarding resolution, approximately 55% of videos are at VGA resolution (854 × 480), 32.5% at HD resolution (1,280 × 720), and 12.5% at full HD resolution (1,920 × 1,080).

Deepfake detection challenge dataset

The Deepfake detection challenge (DFDC) dataset, developed by Facebook (Dolhansky et al., 2020), comprises a curated assemblage of 5,000 face videos that have been modified, showcasing a specific group of performers. The actor selection process was influenced by distinct attributes, resulting in a distribution of 74% female and 26% male actors. Additional classification was conducted according to ethnicity, revealing that 68% of the population identified as Caucasians, 20% as African-American, 9% as West Asian, and 3% as South Asian. Applying face swap techniques resulted in two outcomes: producing high-quality photos of faces close to the camera while maintaining their original proportions and generating lower-quality photographs featuring swapped faces. The dataset consists of 780 testing clips and 4,464 training clips, all 15 s long and have different resolutions.

HOHA-based dataset

The dataset contains eight distinct kinds of human actions that have been extracted from a selection of 32 widely recognized Hollywood films. Priti et al. (2021) presented a dataset comprising 300 randomly selected films obtained from the HOHA dataset. This dataset included 16 samples highlighting human activities extracted from well-known movies and 300 deepfake videos acquired from several internet video platforms. There are 600 videos, each operating at an approximate frame rate of 24 frames per second and with a 360 by 240 pixels resolution.

FakeAVCeleb dataset

Khalid et al. (2021) presents a novel audio-video multimodal deepfake dataset that contains both deepfake videos and their corresponding synthesized cloned audios. This dataset was created using the most popular deepfake generation methods, and the videos and audios are perfectly lip-synced. It aims to address the challenge of impersonation attacks using deepfake videos and audios, with content selected from real YouTube videos of celebrities covering four racial backgrounds to counter the racial bias issue.

Attack agnostic dataset

Kawa, Plata & Syga (2022) combines two audio DeepFakes and one anti-spoofing dataset. This dataset aims to improve the generalization and stabilization of audio DeepFake detection methods by employing a disjoint use of attacks. It is designed to challenge and enhance the detection capabilities of algorithms against various types of spoofing attacks not explicitly included in the training.

ADD dataset

This dataset is part of the first Audio Deep Synthesis Detection (ADD) challenge, which includes tracks for low-quality fake audio detection, partially fake audio detection, and an audio fake game. This dataset by Wu et al. (2023) covers a range of real-life and challenging scenarios, aiming to push the boundaries of audio deepfake detection research.

TweepFake dataset

Fagni et al. (2021) presents TweepFake dataset which is designed to tackle the challenge of detecting deepfake social media messages. It was created in response to the advancements in language modeling, such as the release of GPT-2 by OpenAI, which significantly improved the generative capabilities of deep neural models. These models can autonomously generate coherent, non-trivial, and human-like text samples, posing a threat when used maliciously to generate plausible deepfake messages on social networks like Twitter or Facebook. The dataset is real in the sense that each deepfake tweet was actually posted on Twitter, collected from 23 bots imitating 17 human accounts. The bots are based on various generation techniques, including Markov Chains, RNN, RNN+Markov, LSTM, and GPT-2. To create a balanced dataset, tweets from the humans imitated by the bots were also selected, resulting in a total of 25,572 tweets (half human and half bots generated). The dataset is publicly available on Kaggle.

ADBT dataset

The study by Julien, Babacar & Adrian (2022) focuses on the automatic detection of bot-generated Twitter messages, acknowledging the challenge posed by recent deep language models capable of generating textual content difficult to distinguish from that produced by humans. Such content, potentially used in disinformation campaigns, can have amplified detrimental effects if it spreads on social networks. The study proposes a challenging definition of the problem by making no assumptions regarding the bot account, its network, or the method used to generate the text. Two approaches based on pretrained language models are devised, and a new dataset of generated tweets is created to enhance the classifier’s performance on recent text generation algorithms.

Table 1 provides a comprehensive overview of the key datasets commonly applied for deepfake detection, summarizing their key attributes. Figure 4 organizes the diverse array of datasets used in deepfake detection, providing a taxonomy based on content type—image, video, or combined. This categorization facilitates a deeper understanding of how different data modalities can be utilized or combined to enhance detection. For instance, CelebA is widely used for image-based deepfake training, while FaceForensics++ is a benchmark in video deepfake detection. The representation of datasets highlights the trend towards multi-modal datasets to improve the robustness of detection algorithms against sophisticated deepfakes.

Applied deepfake technology & applications

Deepfakes have rapidly evolved thanks to advances in artificial intelligence and machine learning. The power to generate convincing synthetic media is no longer restricted to professional artists and studios. It has moved into the hands of the common individual, facilitated by deep learning algorithms. This section dissects the techniques employed in deepfake generation and highlights the pivotal role of artificial intelligence in this process.

Figure 5 illustrates the significant increase in the prevalence of deepfakes from 2017 to 2023, alongside improvements in detection accuracy. This trend can be attributed to advancements in deep learning algorithms and the availability of large datasets for training. In 2017, deepfake technology was relatively nascent, with limited applications primarily in academic settings (Khormali & Yuan, 2021). However, by 2023, the technology has evolved rapidly, leading to more sophisticated deepfakes that are harder to detect.In the next few sections we will discuss in details the video, audio, text and image deepfake survey which will help significantly in countering the main problem.

Video deepfake

There has been a notable improvement in detection methods. Early detection systems relied heavily on simple visual and audio cues, which were effective against the initial generation of deepfakes (Lomnitz et al., 2020). As deepfakes became more sophisticated, detection methods evolved to use more complex features, including biometric and behavioral patterns, and employed advanced machine learning techniques, particularly deep neural networks. Masood et al. (2023) indicated that deepfakes in 2023 rose to around 90%. The improvement in detection accuracy is a result of both the evolution of the technology and the growing awareness of the need for robust detection systems to combat the rise in deepfakes. Vera (2022) examines the phenomenon of deepfake generation by analyzing instructional videos about deepfake creation available on the popular video-sharing platform YouTube. The researcher performed a theme analysis on YouTube videos about the construction of deepfakes to comprehend the individuals involved in their production and the underlying motivations driving their creation. The results suggest a greater representation of individuals from non-western backgrounds involved in producing deepfake content. This article examines the influence of deepfakes on how professionals in the field of Library and Information Science (LIS) perceive and address issues related to mis/disinformation. It argues that to combat the negative effects of deepfakes effectively, it is crucial to comprehend their implications within broader socio-technical discourses.

Hameleers (2023) and Yasur et al. (2023) present the notion of misinformation as an intentional and context-dependent action in which individuals surreptitiously deceive recipients by distorting, altering, or creating information to achieve maximum benefit, specifically by misleading the recipients. The conceptual framework thoroughly explains the factors that influence the decision-making process of various individuals or entities to engage in deceptive behavior. It also delves into the strategies employed in deception and the specific objectives that deceivers seek to accomplish by misleading their intended targets. The results of this study have the potential to contribute to the development of machine-learning methods for detecting deception, as well as treatments designed to induce skepticism by challenging the default assumption of truth. Abu-Ein et al. (2022) offers a comprehensive examination of the algorithms and datasets employed in developing deepfakes and an analysis of the existing methods proposed for detecting deepfakes. The authors comprehensively examine the underlying principles and techniques employed in deepfake methodologies. They present a thorough survey of existing deepfake approaches and advocate for developing and implementing novel and resilient ways to effectively address the growing intricacies associated with deepfakes. Gu et al. (2022) extensively examine videos’ localized motion characteristics to identify and detect instances of deepfakes. The authors contend that current methods for detecting deepfake videos fail to consider the local motions between neighboring frames. These motions provide valuable inconsistent information that can be utilized as an effective indicator for detecting deepfake videos. The authors suggest introducing a new sampling unit called a “snippet”. This unit consists of several consecutive video frames to learn local temporal inconsistencies.

The authors have developed an Intra-Snippet Inconsistency Module (Intra-SIM) and an Inter-Snippet Interaction Module (Inter-SIM) to create a framework for dynamic inconsistency modeling. The Intra-SIM algorithm utilizes bi-directional temporal difference operations and a learnable convolution kernel to extract short-term motions within each snippet. The Inter-SIM is designed to facilitate the exchange of information amongst snippets, enabling the construction of global representations. The proposed method demonstrates superior performance compared to existing state-of-the-art rivals across four widely used benchmark datasets, namely FaceForensics++, Celeb-DF, DFDC, and Wild-Deepfake. This study presents a novel approach for identifying deepfakes by emphasizing local motion and temporal inconsistency, which can be seamlessly included in pre-existing 2D convolutional neural networks (CNNs).

The pioneering work by Thies et al. (2016) presented a technique called Face2Face that manipulates video footage in real-time. Using RGB input, the method models the facial geometry and texture, making it possible to transfer the facial expressions of one individual to another in a video. This early deepfake technique reveals the potential for real-time deepfake generation. Güera & Delp (2018) presented a method to detect DeepFake videos using recurrent neural networks (RNNs). The authors proposed an LSTM-based architecture that extracts temporal features from face sequences to identify manipulated videos. The study demonstrated the efficacy of deep learning in not just creating but also detecting deepfakes. Li, Ming-Ching & Siwei (2018) have highlighted a limitation in generating deepfakes—the inability to replicate eye blinking in synthetic videos accurately. They proposed an AI-based approach that uses this cue to detect deepfake videos. This article reminds us that even in the face of advanced deepfake techniques, subtle human elements can be hard to replicate. Hsu, Lee & Zhuang (2018) and Nguyen, Yamagishi & Echizen (2019) have proposed an approach to detecting fake face images in real-world scenarios and presented a novel application of capsule networks, a recent development in deep learning, to detect forged images and videos. They trained a deep learning model to identify manipulated images based on minor inconsistencies typically introduced during the creation of deepfakes, offering a practical tool for mitigating the impact of deepfakes in real-life settings. Their research emphasized the potential of emerging AI techniques in deepfake detection, opening new avenues for future exploration. Durall Lopez et al. (2019) has unveiled a simple yet powerful method to detect deepfake videos based on basic features like eye movement and light reflection. This study demonstrated that even relatively simple features could be potent tools for identifying deepfake content. Khder et al. (2023) examines the significance of artificial intelligence in developing and identifying deepfakes within multimedia. The authors emphasize the emergence of deepfakes, a modified content type made possible by artificial intelligence advancements, which closely mimic authentic videos. This presents a substantial obstacle in accurately discerning between genuine and altered media. This research presents a comprehensive study that systematically evaluates relevant scholarly literature to investigate the fundamental elements and potential remedies associated with deepfakes. The authors’ conclusion posits that the proliferation of deepfakes can amplify risks to nations and digital civic cultures by fostering a pervasive atmosphere of skepticism and uncertainty toward online information sources. This study offers significant insights into the ramifications of deepfakes across many forms of multimedia facilitated by artificial intelligence (AI) tools while underscoring the imperative for developing and implementing robust detection methodologies.

Using frame comparison analysis, the study Ahmad &Shaun (2023) shows how to find deep fakes quickly and accurately. The authors say that even though deepfake technology has improved quickly, the tools for finding deepfakes are often too hard to use and harder to find than the tools for making deepfakes. The suggested method aims to find deepfake videos faster than the CNN-based methods that are already in use. When a deepfake video is made, the edges of the imposed face are handled by distorting the edges. This makes the edges of the face blurry and noisy. This is used by the suggested model, which compares the frames using the Laplacian operator to find edges and deepfake. Pishori et al. (2020) looks at three ways to find deepfake movies. These methods include convolutional LSTM, eye blink recognition, and grayscale histograms. They participated in the Deepfake Detection Challenge, a competition to improve deepfake detection technology by putting out a lot of new data. The first way that was looked at was the convolutional LSTM model. This model uses a CNN to extract frame features and a long short-term memory (LSTM) network to analyze the order of events in time. The model was trained with data from 600 movies, and with only 40 frames from each video, it was accurate more than 97% of the time.

In the study, eye-blink tracking and grayscale histogram analysis, two well-known ways to find deepfake videos, were carefully tested. The first one used a sophisticated long-term recurrent convolutional network (LCRN) to record the time-sequenced interaction of frames that showed eyes opening and closing, and it got 99% right. The second method was based on the small but noticeable differences between the grayscale histograms of movies made by a generative adversarial network (GAN) and those made by regular cameras. Standardizing the histograms ensured that each one had the same effect on the model. The grayscale histogram method did better than the other two, but both were effective between 80% and 90%, with only small improvements over the standard models. Rashid, Lee & Kwon (2021) proposes a novel approach to combat deepfake videos and protect the integrity of videos/images using blockchain technology. The authors argued that blockchain, with its inherent security features such as immutability and transparency, can effectively maintain a secure and tamper-proof record of original videos/images. The proposed technique involves creating a unique digital fingerprint (hash) for each original video/image and storing it in a blockchain. When a video/image is presented, its hash is compared with the one stored in the blockchain. If the hashes match, the video/image is verified as original; otherwise, it is flagged as potentially manipulated or deepfaked.

The strength of this technique lies in the robustness and security of blockchain technology, which makes it nearly impossible for malicious actors to alter the stored hashes. However, the limitation is that the original videos/images must be hashed and stored in the blockchain beforehand, which might only be feasible in some scenarios.

Lastly, Ikram, Chambial & Sood (2023) show a method for detecting deepfake videos that uses a hybrid CNN made up of InceptionResnet v2 and Xception to pull out frame-level features. The DFDC deepfake detection competition dataset on Kaggle was used to train and test the system.

The mixed model got a precision of 0.985, a recall of 0.96, a f1-score of 0.98, and a support of 0.968, which shows that it can find deepfake videos with high accuracy. The strength of this method is that it uses a hybrid CNN model, which blends the strengths of InceptionResnet v2 and Xception to find more robust and differentiable features for detecting deepfakes.

However, the problem with this method is that it needs many training data to work well. Also, while the model worked well on the DFDC dataset, it may not work well on other datasets or in the real world.

Image deepfake

Jevnisek & Avidan (2022) presents a new methodology for detecting deepfakes. This strategy involves the aggregation of features that are taken from all layers of a backbone network. The authors contend that a significant portion of prior research operates under the assumption that the deepfakes present in the test set are generated utilizing the identical strategies employed during the network’s training phase. This assumption diverges from the conditions observed in real-world scenarios. In practical scenarios, training a network using one deepfake algorithm and evaluating its performance on deepfakes created by a different algorithm is common. The algorithm proceeds by implementing a sequential process of extracting visual features from a deep backbone, followed by a binary classification head. Instead of transmitting data throughout the network until it reaches the classification head responsible for determining the authenticity of an image, the algorithm consolidates features derived from all tiers of a single backbone network to identify counterfeit content. This methodology enables the classification head to concurrently utilize many features associated with distinct receptive fields in the image plane.

Suganthi et al. (2022) assess their methodology across two domains: deepfake identification and synthetic picture detection. They demonstrate that their strategy attains outcomes that are considered the most advanced in these areas. Additionally, the authors suggest employing the Coefficient of Variation (CoV) of average precision scores as a metric for evaluating the efficacy of diverse algorithms across multiple datasets.

Additionally, it offers a novel approach to optimizing the performance of a foundational network by adjusting its layers. The system is designed to operate on the detection of deepfake and synthetic images. The model attains a state-of-the-art level of performance in terms of overall accuracy when applied to cross-dataset generalization. This refers to training the model on one deepfake or synthetic image dataset and testing it on a different dataset. Korshunov & Sebastien (2018) have discussed the threat of DeepFakes to face recognition systems. They analyzed the performance of automatic face recognition before and after applying DeepFake manipulations, suggesting that DeepFakes can effectively bypass face recognition systems, thereby raising significant security concerns.

Most importantly, Seibold et al. (2017) have presented a deep learning-based method to detect morphed face images. Their proposed CNN model, MoDe, was trained on a large-scale morphed face dataset. The study demonstrated the potential of using deep learning for detecting sophisticated face morphing attacks, a technique often used in deepfake generation. Zhou et al. (2017) have proposed a two-stream CNN for tampered face detection, where one stream processes facial landmarks while the other handles texture information. The work demonstrated a technique to synergize multiple neural networks for more efficient deepfake detection. Xie, Xu & Ji (2022) have presented a self-supervised learning approach for identifying manipulations in images and videos, underscoring the role of AI in learning robust representations from the data itself without human-annotated labels. This article signals a shift towards self-supervised techniques in the battle against deepfakes.

MesoNet has been presented by Afchar et al. (2018), a compact yet effective model to detect facial forgeries in videos. MesoNet was designed to distinguish between genuine and tampered content, even in challenging scenarios where the quality of the forgery is very high. This study shows the balance between complexity and performance in deepfake detection models. Agarwal et al. (2019) focuses on the threat of deepfakes targeting world leaders. The authors propose a novel method for detecting deepfakes using a combination of traditional image forensics and deep learning techniques. They begin by discussing the potential harm of deepfakes, particularly when used to target world leaders.

The authors then present their proposed method for deepfake detection. They use a two-step approach, first using traditional image forensics to detect anomalies in the images and then a deep learning model to classify the images as real or fake. They also discuss the datasets used for training their model and the performance of their method compared to other deepfake detection techniques. Park, Lim & Kwon (2023) examines the efficacy of facial image alteration techniques within online educational settings, specifically focusing on animating-based facial image manipulation. The authors’ primary focus lies in examining expression swap, a specific technique employed in facial image modification that enables the alteration of facial expressions only within an image. The division of expression swap can be categorized into two distinct approaches: learning-based swap and animating-based swap. The authors assess these strategies’ performance in various scenarios, such as attendance checks, presentations, and tests. The study demonstrates that both methodologies yield satisfactory outcomes when the facial region constitutes a substantial image component. Nevertheless, their performances’ efficacy noticeably diminishes when the facial region constitutes a lesser proportion of the visual content. This study employs face image editing techniques based on animation, specifically focusing on expression swapping. Saravana Ram et al. (2023) presents an innovative approach to deepfake detection using a computer vision-based deep neural network with pairwise learning. The authors propose a novel method that leverages the power of deep learning and computer vision to detect deepfakes effectively. The technique involves using a pairwise learning model that compares pairs of images to determine if they are genuine or manipulated.

The authors argue that traditional deepfake detection methods often fail to detect sophisticated deepfakes due to their inability to capture subtle inconsistencies in manipulated images. Their proposed model uses a deep neural network trained on a large dataset of real and deepfake images to address this. The model learns to identify the subtle differences between real and fake images by comparing pairs of images. The pairwise learning approach is particularly effective in detecting deepfakes as it can capture the subtle inconsistencies in manipulated images that are often overlooked by traditional detection methods. However, the authors acknowledge that the training data quality and the deepfakes’ complexity can affect the model’s performance. Neekhara et al. (2021) presents an in-depth analysis of the malicious threats to deepfake detection. The authors examine the vulnerabilities of deepfake detectors and propose countermeasures to improve their robustness. The authors begin by discussing the adversarial attacks on deepfake detectors. They explain how these attacks exploit the weaknesses of the detectors, such as their reliance on specific features or patterns, to evade detection. The authors also highlight the potential consequences of these attacks, including the spread of misinformation and the erosion of trust in digital media. The article then presents a series of experiments designed to evaluate the robustness of deepfake detectors against adversarial attacks. The authors use various attack techniques, including perturbation-based and poisoning attacks, and test them on several popular deepfake detectors. The results reveal significant vulnerabilities in the tested detectors, with many failing to identify deepfakes under adversarial attacks accurately. These findings underscore the need for more robust, resilient deepfake detection methods. Sun et al. (2022) introduce WildDeepfake, a challenging real-world dataset for deepfake detection. The authors argue that existing datasets for deepfake detection are not representative of real-world scenarios, and thus, they propose WildDeepfake to address this gap. The authors begin by discussing the limitations of existing deepfake datasets. They argue that these datasets often contain synthetic media created under controlled conditions, which may not accurately reflect the diversity and complexity of real-world deepfakes.

To address this issue, WildDeepfake is introduced, a dataset comprised of deepfakes collected from various online sources. The dataset includes a wide range of deepfakes, from amateur creations to sophisticated productions, providing a more realistic benchmark for deepfake detection methods. A series of experiments are done to evaluate the performance of several popular deepfake detectors on the WildDeepfake dataset. The results reveal that these detectors struggle to accurately identify deepfakes in the dataset, underscoring the challenges of real-world deepfake detection.

Digital forensic methods for deepfake detection in social media

The advent of deepfake technology has necessitated the development of digital forensic methods to detect and analyze these artificially synthesized media. Deepfakes, which convincingly replace parts of original content with fabricated elements, pose significant threats ranging from personal identity theft to the propagation of disinformation (Güera & Delp, 2018).

Digital forensic methods are the first defense against these threats, providing computational techniques to identify and analyze deepfakes. The evolution of deepfake detection methods has been driven by the increasing sophistication of deepfake generation techniques. Initial detection methods relied on manual or rudimentary automated techniques, focusing on inconsistencies in lighting, blurring, and artifacts from image compression (Maras & Alexandrou, 2019).

However, the rise of deep learning and artificial intelligence has revolutionized the field, leading to the development of more advanced and accurate detection methods. These modern methods leverage machine learning algorithms to analyze intricate patterns and subtle inconsistencies typically invisible to the human eye (Nguyen, Yamagishi & Echizen, 2019).

The effectiveness of digital forensic methods in detecting deepfakes can vary across different modalities, such as images, audio, and video. Moreover, their applicability is continually challenged by the dynamic landscape of social media, where the rapid dissemination of content and the evolving nature of deepfakes necessitate constant advancements in detection techniques. Therefore, ongoing research in this field is crucial to keep pace with the evolving threats and ensure digital media’s integrity and trustworthiness (Wang et al., 2022a; Wang et al., 2022b; Wang et al., 2022c).

This section delves into the various digital forensic methods developed and employed to detect and analyze deepfakes. It comprehensively evaluates these methods, assessing their effectiveness in detecting deepfakes across different modalities such as images, audio, and video.

The section also discusses the applicability of these methods within the dynamic and rapidly evolving landscape of social media, highlighting the challenges and opportunities in this context. The aim is to provide a thorough understanding of the current state of deepfake detection techniques and their performance in real-world scenarios, particularly in social media platforms where the dissemination of deepfakes can have significant impacts.

Table 2 summarizes the method used, modality, features, datasets, performance metrics, and advantages and limitations of the techniques mentioned above.

Emma Watson deepfake incident

The world of deepfakes felt its impact when renowned actress Emma Watson, famous for her role in the Harry Potter series, became a victim of this technology. Her face was superimposed onto explicit content, causing a significant stir as the public was still largely unfamiliar with deepfakes. The incident profoundly affected her reputation, demonstrating the potential harm deepfakes can cause individuals.

Natalie Portman deepfake scandal

Similar to Emma Watson, Natalie Portman, another acclaimed actress, was targeted by deepfake creators. Her likeness was used in explicit videos, leading to widespread shock and confusion among fans and the public. This incident underscored the ease with which deepfakes can be used to tarnish reputations.

Cheerleader deepfake case

In a disturbing turn of events, a mother was accused of creating deepfake videos of her daughter’s cheerleading teammates, depicting them in compromising situations to intimidate them into leaving the team. However, the lack of reliable detection tools led to the case being dismissed, highlighting the challenges in proving the use of deepfakes.

Zelensky deepfake incident

Amidst the Russia-Ukraine conflict, a deepfake video of Ukrainian President Zelensky surfaced, in which he appeared to be urging Ukrainians to surrender. This incident caused temporary panic and demonstrated the potential of deepfakes to influence political situations.

Kobe Bryant deepfake in music video

The late NBA star Kobe Bryant was brought back to life in a Kendrick Lamar music video using deepfake technology. The realistic portrayal resonated with fans, showing the potential of deepfakes in the entertainment industry.

Tom cruise tiktok deepfake

A deepfake of Tom Cruise, created by VFX artist Chris Ume and TikTok user Miles Fisher, gained significant attention due to its high level of detail, demonstrating the sophistication of current deepfake technology.

Jim Carrey in the Shining deepfake

A fan-made deepfake inserted comedian Jim Carrey into a scene from The Shining, replacing Jack Nicholson. The deepfake’s high quality and Carrey’s uncanny impersonation of Nicholson highlighted the potential of deepfakes in film.

Barack Obama deepfake PSA

A deepfake of former President Barack Obama was created by BuzzFeedVideo to raise awareness about the potential misuse of deepfake technology. The video, featuring actor Jordan Peele impersonating Obama, emphasized the need for responsible internet usage.

Donald Trump reindeer story deepfake

A comedic deepfake of former President Donald Trump was uploaded by the YouTube channel Sassy Justice, showing Trump telling a nonsensical story about a reindeer. Despite some audio inaccuracies, the video demonstrated the potential for deepfakes in satire and comedy.

Hillary Clinton SNL deepfake

A deepfake of former Secretary of State Hillary Clinton was used in an episode of SNL, with a cast member impersonating her. The deepfake’s high quality showed the increasing sophistication of this technology, emphasizing the need for caution.

Kim Joo-Ha deepfake incident

In a groundbreaking move, South Korean television channel MBN utilized deepfake technology to replace popular news anchor Kim Joo-Ha with an AI version of herself. The AI, bearing the likeness and voice of Kim Joo-Ha, was used for various news reports, including morning news and traffic updates. This incident raises questions about the potential implications of deepfakes on the media industry and the future of professions like newsreading.

Lynda Carter as Wonder Woman deepfake

Deepfake technology breathed new life into beloved characters from the past, as demonstrated by the deepfake of Lynda Carter, the original Wonder Woman from the 1970s TV show, replacing Gal Gadot in the recent Wonder Woman films. This realistic deepfake hints at the potential changes deepfake technology could bring to the entertainment industry.

Luke Skywalker deepfake in the Mandalorian

In the Mandalorian series, the Star Wars franchise saw the return of a beloved character, Luke Skywalker. However, the CGI recreation of Mark Hamill, the original actor, had flaws. A deepfake version of the character showcased a higher quality rendition, demonstrating the potential superiority of deepfake technology over traditional CGI.

Donald Trump in Better Call Saul deepfake

Deepfakes can also serve comedic purposes, as seen in a parody where former President Donald Trump was substituted for Saul Goodman, a character from the popular series Breaking Bad and its spin-off, Better Call Saul. This humorous deepfake underscores the potential of deepfakes in creating entertaining content.

Mark Zuckerberg deepfake

In response to Facebook’s refusal to delete an edited video of Nancy Pelosi, artist Bill Posters created a deepfake of Mark Zuckerberg claiming ownership over Facebook’s users. While the deepfake was not entirely convincing due to a voice mismatch, it served as a statement on the potential misuse of deepfake technology in manipulating public opinion.

Bill Hader deepfake

Known for his impressive impersonations, comedian Bill Hader was the subject of a deepfake where his face was overlaid with those of Al Pacino and Arnold Schwarzenegger during his impersonations. While not entirely perfect, the deepfake, combined with Hader’s voice and mannerisms, provided an entertaining and harmless example of deepfake technology.

The ever-evolving nature of deepfake technology

Deepfake technology is advancing at an unprecedented rate, with new methods and techniques emerging regularly. This rapid evolution makes it difficult for digital forensic techniques to keep pace. For instance, the development of AI-based techniques for synthesizing human images, known as Deep Fakes, has made it possible to create highly realistic fake videos and images that leave few traces of manipulation (Liu et al., 2023).

Moreover, deepfake technology is improving in terms of quality but also terms of accessibility and speed. Today, deepfakes can be generated in real-time, enabling attackers to impersonate people over audio and video calls (Almutairi & Elgibreen, 2022). This real-time capability significantly increases the potential for misuse of deepfake technology and poses new challenges for detection methods.

Limitations of current digital forensic techniques

Current digital forensic techniques face several limitations in detecting deepfakes. One of the primary challenges is the generalizability of detection methods across different deepfake generation schemes (Wang et al., 2022a; Wang et al., 2022b; Wang et al., 2022c). A reliable deepfake detection approach must be agnostic to the type of deepfake, which can present diverse quality and appearance. However, many existing detection methods fail to handle unseen attacks in an open set scenario, limiting their effectiveness (Xu et al., 2023).

Another significant challenge is the increasing sophistication of deepfake technology. With advanced AI and machine learning techniques, deepfakes are becoming more realistic and harder to detect. For instance, hybrid CNN has been shown to increase the accuracy of deepfake generation, making it more challenging for detection methods to identify fake content (Ikram, Chambial & Sood, 2023).

Furthermore, the effectiveness of detection methods is also limited by the quality and quantity of data available for training. The performance of deepfake detection methods often relies on large, high-quality datasets for training. However, the availability of such datasets is limited, and the rapid evolution of deepfake technology can quickly render existing datasets outdated (Hussain & Ibraheem, 2023). The challenges posed by deepfake technology are not limited to the technical aspects of detection. They also extend to the societal implications of these manipulations. As deepfake technology becomes more sophisticated, it is increasingly used for malicious purposes, such as spreading disinformation, committing fraud, and violating privacy. This has led to growing concerns about the potential misuse of this technology and the need for effective countermeasures. One of the major challenges in combating deepfakes is the rapid pace of advancement in deepfake generation techniques. As highlighted by Akhtar (2023), there is a continuous arms race between the creators of deepfake generation methods and the developers of detection techniques. It provides an overview of the various deepfake generation and detection methods and discusses the open challenges and potential research directions in this field (Siegel et al., 2021; Lan et al., 2022).

Another challenge is the degradation of the performance of deepfake detection systems under various manipulations. This is discussed by Liu et al. (2023), which presents the results of the ASVspoof 2021 challenge. While detection systems offer some resilience to compression effects, they lack generalization across different source datasets2. Whereas Almutairi & Elgibreen (2022) emphasize a need to review existing audio deepfake detection methods comprehensively. The need for more robust detection models to detect fakeness even in the presence of accented voices or real-world noises is highlighted. While significant progress has been made in detecting deepfakes, many challenges remain to overcome. The rapid evolution of deepfake technology, the societal implications of deepfakes, and the limitations of current detection techniques all contribute to the complexity of this issue. Continuous research and innovation are needed to keep pace with these advancements and to develop more effective countermeasures against deepfakes (Abdulreda & Obaid, 2022a; Abdulreda & Obaid, 2022b).