Deepfake Audio And Video Detection For Digital Forensics — Complete Documentation & Simulation

Overview & System Context

Deepfake Audio And Video Detection For Digital Forensics: Comprehensive Web-Sourced Technical Investigation Exhaustive Research Dossier · Pure Web-Scraped Content: ~9546 Words · Real Software Screenshots: 3 1. Executive Overview & Technical Background Audio deepfake technology, also referred to as voice cloning or deepfake audio, is an application of artificial intelligence designed to generate speech that convincingly mimics specific individuals, often synthesizing phrases or sentences they have never spoken. Initially developed with the intent to enhance various aspects of human life, it has practical applications such as generating audiobooks and assisting individuals who have lost their voices due to medical conditions. Additionally, it has commercial uses, including the creation of personalized digital assistants, natural-sounding text-to-speech systems, and advanced speech translation services. As with other forms of deepfakes, audio deepfakes have faced criticism for their use in scams, phishing, and misinformation campaigns.

The category based on speech synthesis refers to the artificial production of human speech, using software or hardware system programs. Speech synthesis includes text-to-speech, which aims to transform the text into acceptable and natural speech in real-time, making the speech sound in line with the text input, using the rules of linguistic description of the text. A classical system of this type consists of three modules: a text analysis model, an acoustic model, and a vocoder. The generation usually has to follow two essential steps. It is necessary to collect clean and well-structured raw audio with the transcribed text of the original speech audio sentence. Second, the text-to-speech model must be trained using these data to build a synthetic audio generation model. Specifically, the transcribed text with the target speaker's voice is the input of the generation model. The text analysis module processes the input text and converts it into linguistic features. Then, the acoustic module extracts the parameters of the target speaker from the audio data based on the linguistic features generated by the text analysis module. Finally, the vocoder learns to create vocal waveforms based on the parameters of the acoustic features.

The final audio file is generated, including the synthetic simulation audio in a waveform format, creating speech audio in the voice of many speakers, even those not in training. The first breakthrough in this regard was introduced by WaveNet, a neural network for generating raw audio waveforms capable of emulating the characteristics of many different speakers. This network has been overtaken over the years by other systems which synthesize highly realistic artificial voices within everyone's reach. Text-to-speech is highly dependent on the quality of the voice corpus used to realize the system, and creating an entire voice corpus is expensive. Another disadvantage is that speech synthesis systems do not recognize periods or special characters. Also, ambiguity problems are persistent, as two words written in the same way can have different meanings.

Audio deepfake based on imitation is a way of transforming an original speech from one speaker - the original - so that it sounds spoken like another speaker - the target one. An imitation-based algorithm takes a spoken signal as input and alters it by changing its style, intonation, or prosody, trying to mimic the target voice without changing the linguistic information. This technique is also known as voice conversion. This method is often confused with the previous synthetic-based method, as there is no clear separation between the two approaches regarding the generation process. Indeed, both methods modify acoustic-spectral and style characteristics of the speech audio signal, but the Imitation-based usually keeps the input and output text unaltered. This is obtained by changing how this sentence is spoken to match the target speaker's characteristics.

Voices can be imitated in several ways, such as using humans with similar voices that can mimic the original speaker. In recent years, the most popular approach involves the use of particular neural networks called generative adversarial networks (GAN) due to their flexibility as well as high-quality results. Then, the original audio signal is transformed to say a speech in the target audio using an imitation generation method that generates a new speech, shown in the fake one.

Many machine learning models have been developed using different strategies to detect fake audio. Most of the time, these algorithms follow a three-steps procedure: Figure 1: UK Deepfake Detection Framework: Microsoft and Government Unite - AI ... Each speech audio recording must be preprocessed and transformed into appropriate audio features; The computed features are fed into the detection model, which performs the necessary operations, such as the training process, essential to discriminate between real and fake speech audio; The output is fed into the final module to produce a prediction probability of the Fake class or the Real one. Following the ASVspoof challenge nomenclature, the Fake audio is indicated with the term "Spoof," the Real instead is called "Bonafide." Over the years, many researchers have shown that machine learning approaches are more accurate than deep learning methods, regardless of the features used. However, the scalability of machine learning methods is not confirmed due to excessive training and manual feature extraction, especially with many audio files. Instead, when deep learning algorithms are used, specific transformations are required on the audio files to ensure that the algorithms can handle them. Most detection studies focus on English language deepfakes. Research on the most spoken languages like Chinese and Spanish remains limited. Hindi and Arabic are also understudied. Detection models trained primarily on English show reduced performance when tested on non-English audio. Accent variation also affects detection accuracy. Different pronunciations associated with specific regions or populations influence how well models identify synthetic speech. There are several open-source implementations of different detection methods, and usually many research groups release them on a public hosting service like GitHub.

The increasing accessibility of audio deepfake technology has faced concerns over their use in scams and misinformation campaigns. People can use them as a logical access voice spoofing technique, where they can be used to manipulate public opinion for propaganda, defamation, or terrorism. Vast amounts of voice recordings are daily transmitted over the Internet, and spoofing detection is challenging. Audio deepfake attackers have targeted individuals and organizations, including politicians and governments. In 2019, scammers using AI impersonated the voice of the CEO of a German energy company and directed the CEO of its UK subsidiary to transfer €220,000. In early 2020, the same technique impersonated a company director as part of an elaborate scheme that convinced a branch manager to transfer $35 million. In May 2024, the British engineering firm Arup confirmed that it had been the target of a deepfake scam in Hong Kong in which fake voices and images were used. According to Hong Kong police, a finance employee joined a video call with deepfake re-creations of the company's chief financial officer and other colleagues, and then transferred 200 million Hong Kong dollars (about US$25.6 million) to the fraudsters in 15 transactions. According to a 2023 global McAfee survey, one person in ten reported having been targeted by an AI voice cloning scam; 77% of these targets reported losing money to the scam. Audio deepfakes could also pose a danger to voice ID systems currently used by financial institutions. In March 2023, the United States Federal Trade Commission issued a warning to consumers about the use of AI to fake the voice of a family member in distress asking for money. In October 2023, during the start of the British Labour Party's conference in Liverpool, an audio deepfake of Labour leader Keir Starmer was released that falsely portrayed him verbally abusing his staffers and criticizing Liverpool. That same month, an audio deepfake of Slovak politician Michal ŠimeIka falsely claimed to capture him discussing ways to rig the upcoming election. During the campaign for the 2024 New Hampshire Democratic presidential primary, over 20,000 voters received robocalls from an AI-impersonated President Joe Biden urging them not to vote. The New Hampshire attorney general said this violated state election laws, and alleged involvement by Life Corporation and Lingo Telecom. In February 2024, the United States Federal Communications Commission banned the use of AI to fake voices in robocalls. That same month, political consultant Steve Kramer admitted that he had commissioned the calls for $500. He said that he wanted to call attention to the need for rules governing the use of AI in political campaigns. In May, the FCC said that Kramer had violated federal law by spoofing the number of a local political figure, and proposed a fine of $6 million. Four New Hampshire counties indicted Kramer on felony counts of voter suppression, and impersonating a candidate, a misdemeanor. In 2024, YouTuber Jeff Geerling received attention after noticing his voice was cloned without consent by the electronics manufacturer Elecrow, and the company subsequently apologized and removed the offending content. Geerling's experience was cited as a case meant to be addressed by YouTube's likeness-detection technology when it was released in 2025.

On 22 May 2025, it was claimed that Hoya Corporations product ReadSpeak used recording work done by the actress Gayanne Potter for them in 2021 which at the time she understood would just be used for accessibility and e-learning software, but is now available generally as the voice Iona and is used as the announcer on ScotRail trains. This replaced older messages recorded by Fletcher Mathers without her permission. On 25 August 2025, ScotRail announced that they will replace the AI voice on trains, however it's not confirmed if this will be a human recording or another AI-trained voice. In September 2026, the Tokyo District Court ruled that the unauthorized commercial use of an AI-generated imitation of a person's voice can infringe publicity rights. It is the first judicial recognition of a right to one's voice in Japan. Voice actor Kenjiro Tsuda brought the case against TikTok over an AI clone of his voice used in more than 180 TikTok videos.

Deepfakes (a portmanteau of 'deep learning' and 'fake') are images, videos, or audio that have been edited or generated using artificial intelligence, AI-based tools or audio-video editing software. They may depict real or fictional people and are considered a form of synthetic media, that is media that is usually created by artificial intelligence systems by combining various media elements into a new media artifact. While the act of creating fake content is not new, deepfakes uniquely leverage machine learning and artificial intelligence techniques, including facial recognition algorithms and artificial neural networks such as variational autoencoders and generative adversarial networks (GANs). In turn, the field of image forensics has worked to develop techniques to detect manipulated images. Deepfakes have garnered widespread attention for their potential use in creating child sexual abuse material, celebrity pornographic videos, revenge porn, fake news, hoaxes, bullying, and financial fraud. Academics have raised concerns about the potential for deepfakes to promote disinformation and hate speech, as well as interfere with elections. In response, the information technology industry and governments have proposed recommendations and methods to detect and mitigate their use. Academic research has also delved deeper into the factors driving deepfake engagement online as well as potential countermeasures to malicious application of deepfakes. From traditional entertainment to gaming, deepfake technology has evolved to be increasingly convincing and available to the public, allowing for the disruption of the entertainment and media industries.

Figure 2: Audio Deepfakes in Digital Forensics: Detection, Risks & Evidence Integrity Generalization. High-quality deepfakes are often achieved by training on hours of footage of the target. This challenge is to minimize the amount of training data and the time to train the model required to produce quality images and to enable the execution of trained models on new identities (unseen during training). Paired Training. Training a supervised model can produce high-quality results, but requires data pairing. This is the process of finding examples of inputs and their desired outputs for the model to learn from. Data pairing is laborious and impractical when training on multiple identities and facial behaviors. Some solutions include self-supervised training (using frames from the same video), the use of unpaired networks such as Cycle-GAN, or the manipulation of network embeddings. Identity leakage. This is where the identity of the driver (i.e., the actor controlling the face in a reenactment) is partially transferred to the generated face. Some solutions proposed include attention mechanisms, few-shot learning, disentanglement, boundary conversions, and skip connections. Occlusions. When part of the face is obstructed with a hand, hair, glasses, or any other item then artifacts can occur. A common occlusion is a closed mouth which hides the inside of the mouth and the teeth. Some solutions include image segmentation during training and in-painting. Temporal coherence. In videos containing deepfakes, artifacts such as flickering and jitter can occur because the network has no context of the preceding frames. Some researchers provide this context or use novel temporal coherence losses to help improve realism. As the technology improves, the interference is diminishing. Overall, deepfakes are expected to have several implications in media and society, media production, media representations, media audiences, gender, law, and regulation, and politics.

On 8 June 2022, Daniel Emmet, a former AGT contestant, teamed up with the AI startup Metaphysic AI, to create a hyperrealistic deepfake to make it appear as Simon Cowell. Cowell, notoriously known for severely critiquing contestants, was on stage interpreting "You're The Inspiration" by Chicago. Emmet sang on stage as an image of Simon Cowell emerged on the screen behind him in flawless synchronicity. On 30 August 2022, Metaphysic AI had 'deep-fake' Simon Cowell, Howie Mandel and Terry Crews singing opera on stage. On 13 September 2022, Metaphysic AI performed with a synthetic version of Elvis Presley for the finals of America's Got Talent. The MIT artificial intelligence project 15.ai has been used for content creation for multiple Internet fandoms, particularly on social media. In 2023 the bands ABBA and Kiss partnered with Industrial Light & Magic and Pophouse Entertainment to develop deepfake avatars capable of performing virtual concerts.

In February 2018, in separate videos, the face of the Argentine President Mauricio Macri had been replaced by the face of Adolf Hitler, and Angela Merkel's face has been replaced with Donald Trump's. In April 2018, Jordan Peele collaborated with BuzzFeed to create a deepfake of Barack Obama with Peele's voice; it served as a public service announcement to increase awareness of deepfakes. In January 2019, Fox affiliate KCPQ aired a deepfake of Trump during his Oval Office address, mocking his appearance and skin colour. The employee found responsible for the video was subsequently fired. In June 2019, the United States House Intelligence Committee held hearings on the potential malicious use of deepfakes to sway elections.

In April 2020, the Belgian branch of Extinction Rebellion published a deepfake video of Belgian Prime Minister Sophie Wilmès on Facebook. The video promoted a possible link between deforestation and COVID-19. It had more than 100,000 views within 24 hours and received many comments. On the Facebook page where the video appeared, many users interpreted the deepfake video as genuine. During the 2020 US presidential campaign, many deepfakes surfaced purporting Joe Biden in cognitive decline—falling asleep during an interview, getting lost, and misspeaking—all bolstering rumours of his decline.

During the 2020 Delhi Legislative Assembly election campaign, the Delhi Bharatiya Janata Party used similar technology to distribute a version of an English-language campaign advertisement by its leader, Manoj Tiwari, translated into Haryanvi to target Haryana voters. A voiceover was provided by an actor, and AI trained using video of Tiwari speeches was used to lip-sync the video to the new voiceover. A party staff member described it as a "positive" use of deepfake technology, which allowed them to "convincingly approach the target audience even if the candidate didn't speak the language of the voter." In 2020, Bruno Sartori produced deepfakes parodying politicians like Jair Bolsonaro and Donald Trump. In April 2021, politicians in a number of European countries were approached by pranksters Vovan and Lexus, who are accused by critics of working for the Russian state. They impersonated Leonid Volkov, a Russian opposition politician and chief of staff of the Russian opposition leader Alexei Navalny's campaign, allegedly through deepfake technology. However, the pair told The Verge that they did not use deepfakes, and just used a look-alike. In May 2023, a deepfake video of Vice President Kamala Harris supposedly slurring her words and speaking nonsensically about today, tomorrow and yesterday went viral on social media. In June 2023, in the United States, Ron DeSantis's presidential campaign used a deepfake to misrepresent Donald Trump. In November 2023, a deepfake video of the German Chancellor Olaf Scholz announcing a plan to ban the political activities of the AfD was uploaded to YouTube by the Zentrum für Politische Schönheit (Center of Political Beauty). In March 2024, during India's state assembly elections, deepfake technology was widely employed by political candidates to reach out to voters.

Many politicians used AI-generated deepfakes created by startup The Indian Deepfaker, founded by Divyendra Singh Jadoun, to translate their speeches into multiple regional languages, allowing them to engage with diverse linguistic communities across the country. This surge in the use of deepfakes for political campaigns marked a significant shift in electioneering tactics in India. In June 2025, Javier Milei's government backed a smear campaign against journalist Mengolini, which was partly based on explicit deepfakes. In July 2025, Donald Trump posted a deepfake on his Truth Social account, depicting former president Barack Obama getting arrested at the White House and put in prison.

In 2017, Deepfake pornography prominently surfaced on the Internet, particularly on Reddit. As of 2019, many deepfakes on the internet feature pornography of female celebrities whose likeness is typically used without their consent. A report published in October 2019 by Dutch cybersecurity startup Deeptrace estimated that 96% of all deepfakes online were pornographic. As of 2018, a Daisy Ridley deepfake first captured attention, among others. As of October 2019, most of the deepfake subjects on the internet were British and American actors. However, around a quarter of the subjects are South Korean, the majority of which are K-pop stars. In June 2019, a downloadable Windows and Linux application called DeepNude was released that used neural networks, specifically generative adversarial networks, to remove clothing from images of women. The app had both a paid and unpaid version, the paid version costing $50. On 27 June the creators removed the application and refunded consumers. Female celebrities are often a main target when it comes to deepfake pornography. In 2023, deepfake porn videos appeared online of Emma Watson and Scarlett Johansson in a face swapping app. In 2024, deepfake porn images circulated online of Taylor Swift. Academic studies have reported that women, LGBT people and people of colour (particularly activists, politicians and those questioning power) are at higher risk of being targets of promulgation of deepfake pornography. Deepfake technology has become a tool for gender-based harassment and violence, proportionally targeting women and marginalized groups. There is an increasing ethical and equity concerns of intimidation and reputational harm intentions behind this specific media.

Figure 3: 12 Deepfake detection tools You Should Know | AI Video Detector The liar's dividend is a political and social phenomenon in which, when faced with incriminating or embarrassing authentic video or audio recordings of themselves, individuals will claim that the recordings are AI in order to dismiss the claims and gain sympathy.

Barack Obama On 17 April 2018, American actor Jordan Peele, BuzzFeed, and Monkeypaw Productions posted a deepfake of Barack Obama to YouTube, which depicted Barack Obama cursing and calling Donald Trump names. In this deepfake, Peele's voice and face were transformed and manipulated into those of Obama. The intent of this video was to portray the dangerous consequences and power of deepfakes, and how deepfakes can make anyone say anything. Donald Trump On 5 May 2019, Derpfakes posted a deepfake of Donald Trump to YouTube, based on a skit Jimmy Fallon performed on The Tonight Show. In the original skit (aired 4 May 2016), Jimmy Fallon dressed as Donald Trump and pretended to participate in a phone call with Barack Obama, conversing in a manner that presented him to be bragging about his primary win in Indiana.

In the deepfake, Jimmy Fallon's face was transformed into Donald Trump's face, with the audio remaining the same. This deepfake video was produced by Derpfakes with a comedic intent. In March 2023, a series of images appeared to show New York Police Department officers restraining Trump. The images, created using Midjourney, were initially posted on Twitter by Eliot Higgins but were later re-shared without context, leading some viewers to believe they were real photographs. Nancy Pelosi In 2019, a clip from Nancy Pelosi's speech at the Center for American Progress (given on 22 May 2019) in which the video was slowed down, in addition to the pitch of the audio being altered, to make it seem as if she were drunk, was widely distributed on social media. Critics argue that this was not a deepfake, but a shallowfakeI—Ia less sophisticated form of video manipulation. Mark Zuckerberg In May 2019, two artists collaborating with the company CannyAI created a deepfake video of Facebook founder Mark Zuckerberg talking about harvesting and controlling data from billions of people. The video was part of an exhibit to educate the public about the dangers of artificial intelligence. Kim Jong-un and Vladimir Putin On 29 September 2020, deepfakes of North Korean leader Kim Jong-un and Russian President Vladimir Putin were uploaded to YouTube, created by a nonpartisan advocacy group RepresentUs. The deepfakes of Kim and Putin were meant to air publicly as commercials to relay the notion that interference by these leaders in US elections would be detrimental to the United States' democracy. The commercials also aimed to shock Americans to realize how fragile democracy is, and how media and news can significantly influence the country's path regardless of credibility. However, while the commercials included an ending comment detailing that the footage was not real, they ultimately did not air due to fears and sensitivity regarding how Americans may react. On 5 June 2023, an unknown source broadcast a reported deepfake of Vladimir Putin on multiple radio and television networks. In the clip, Putin appears to deliver a speech announcing the invasion of Russia and calling for a general mobilization of the army. Volodymyr Zelenskyy On 16 March 2022, a one-minute long deepfake video depicting Ukraine's president Volodymyr Zelenskyy seemingly telling his soldiers to lay down their arms and surrender during the 2022 Russian invasion of Ukraine was circulated on social media. Russian social media boosted it, but after it was debunked, Facebook and YouTube removed it. Twitter allowed the video in tweets where it was exposed as a fake, but said it would be taken down if posted to deceive people. Hackers inserted the disinformation into a live scrolling-text news crawl on TV station Ukraine 24, and the video appeared briefly on the station's website in addition to false claims that Zelenskyy had fled his country's capital, Kyiv. It was not immediately clear who created the deepfake, to which Zelenskyy responded with his own video, saying, "We don't plan to lay down any arms. Until our victory." Wolf News In late 2022, pro-China propagandists started spreading deepfake videos purporting to be from "Wolf News" that used synthetic actors. The technology was developed by a London company called Synthesia, which markets it as a cheap alternative to live actors for training and HR videos. Pope Francis In March 2023, an anonymous construction worker from Chicago used Midjourney to create a fake image of Pope Francis in a white Balenciaga puffer jacket. The image went viral, receiving over twenty million views. Writer Ryan Broderick dubbed it "the first real mass-level AI misinformation case". Experts consulted by Slate characterized the image as unsophisticated: "you could have made it on Photoshop five years ago". Keir Starmer In October 2023, a deepfake audio clip of the UK Labour Party leader Keir Starmer abusing staffers was released on the first day of a Labour Party conference. The clip purported to be an audio tape of Starmer abusing his staffers. Rashmika Mandanna In early November 2023, a famous South Indian actor, Rashmika Mandanna fell prey to DeepFake when a morphed video of a famous British-Indian influencer, Zara Patel, with Rashmika's face started to float on social media. Zara Patel claims to not be involved in its creation. Bongbong Marcos In April 2024, a deepfake video misrepresenting Philippine President Bongbong Marcos was released. It is a slideshow accompanied by a deepfake audio of Marcos purportedly ordering the Armed Forces of the Philippines and special task force to act "however appropriate" should China attack the Philippines. The video was released amidst tensions related to the South China Sea dispute. The Presidential Communications Office has said that there is no such directive from the president and said a foreign actor might be behind the fabricated media. Criminal charges has been filed by the Kapisanan ng mga Brodkaster ng Pilipinas in relation to the deepfake media. On 22 July 2024, a video of Marcos purportedly snorting illegal drugs was released by Claire Contreras, a former supporter of Marcos. Dubbed as the polvoron video, the media noted its consistency with the insinuation of Marcos' predecessor—Rodrigo Duterte—that Marcos is a drug addict; the video was also shown at a Hakbang ng Maisug rally organized by people aligned with Duterte. Two days later, the Philippine National Police and the National Bureau of Investigation, based on their own findings, concluded that the video was created using AI; they further pointed out inconsistencies with the person on the video with Marcos, such as details on the two people's ears. Joe Biden Prior to the 2024 United States presidential election, phone calls imitating the voice of the incumbent Joe Biden were made to dissuade people from voting for him. The person responsible for the calls was charged with voter suppression and impersonating a candidate. The FCC proposed to fine him US$6 million and Lingo Telecom, the company that allegedly relayed the calls, $2 million. Arup Group The firm Arup Group lost $25 million in 2024 from a deepfake scam.

Sara Duterte Philippine Senator Ronald dela Rosa on 14 June 2025 shared a 40-second Deepfake video created via Veo where two students on the street testified on why the impeachment proceedings against Vice President Sara Duterte is selective and politically motivated. Dela Rosa lauded the supposed youth for their opinions. When informed it was an AI video, he says that if that was the case the video creator had a point. Duterte herself defended Dela Rosa, insisting that sharing an AI video is not wrong as long as its "not for profit". Catherine Connolly In October 2025, Irish presidential candidate Catherine Connolly condemned an AI-generated deepfake video that mimicked an RTÉ News report and falsely claimed she was withdrawing from the presidential election; her campaign reported the video to the social media platforms it appeared on.

Sasse, Ben (19 October 2018). "This New Technology Could Send American Politics into a Tailspin". Opinions. The Washington Post. Retrieved 10 July 2019. Fake/Spoof Audio Detection Challenge (ASVspoof) Deepfake Detection Challenge (DFDC) Bibliography: Media Literacy in the Age of Deepfakes. Curated by Dr Joshua Glick. Before you believe – how to recognize a deepfake and is it inherently evil? Tyagi, Vipin (2026). Digital Image Forgery Detection: Techniques, Challenges and Applications.

SpringerBriefs in Computer Science.

Singapore: Springer Nature Singapore.

Deepfake pornography is generative AI pornography created by altering existing photographs or videos, using deepfake technology, to modify the appearance of the depicted individuals, typically without consent. Deepfake pornography is controversial and has been the subject of lawsuits and criminal investigations because users create and distribute realistic photos and videos of non-consenting individuals, sometimes including minors. It has been used for revenge porn. Many countries have criminalized deepfake pornography through legislative measures and technological solutions.

In January 2024, AI-generated sexually explicit images of American singer Taylor Swift were posted on X (formerly Twitter), and spread to other platforms such as Facebook, Reddit and Instagram. One tweet with the images was viewed over 45 million times before being removed. A report from 404 Media found that the images appeared to have originated from a Telegram group, whose members used tools such as Microsoft Designer to generate the images, using misspellings and keyword hacks to work around Designer's content filters. After the material was posted, Swift's fans posted concert footage and images to bury the deepfake images, and reported the accounts posting the deepfakes. Searches for Swift's name were temporarily disabled on X, returning an error message instead. Graphika, a disinformation research firm, traced the creation of the images back to a 4chan community. A source close to Swift told the Daily Mail that she would be considering legal action, saying, "Whether or not legal action will be taken is being decided, but there is one thing that is clear: These fake AI-generated images are abusive, offensive, exploitative, and done without Taylor's consent and/or knowledge." The controversy drew condemnation from White House Press Secretary Karine Jean-Pierre, Microsoft CEO Satya Nadella, the Rape, Abuse & Incest National Network, and SAG-AFTRA. Several US politicians called for federal legislation against deepfake pornography. Later in the month, US senators Dick Durbin, Lindsey Graham, Amy Klobuchar and Josh Hawley introduced a bipartisan bill that would allow victims to sue individuals who produced or possessed "digital forgeries" with intent to distribute, or those who received the material knowing it was made non-consensually.

In December 2025, Bloomberg reported that X users found Grok would comply with non-consensual requests to digitally undress individuals, including minors, or show them performing sexually explicit acts. The majority of these prompts were targeted at women and girls. An analysis of 20,000 images generated by Grok between December 25, 2025 and January 1, 2026 showed 2% were of people in bikinis or transparent clothes and appeared to be 18 or younger, including 30 of "young or very young" women or girls. A separate analysis conducted over 24 hours from January 5 to 6 calculated that users had Grok create 6,700 sexually suggestive or nudified images per hour. xAI responded to requests for comment from media organizations with the automated reply, "Legacy Media Lies". The bot's image generation sparked an international backlash and calls for legal or regulatory action from officials in the European Union, United Kingdom, Poland, France, India, Malaysia, and Brazil.

While both deepfake pornography and generative AI pornography utilize synthetic media, they differ in approach and ethical implications. Generative AI pornography is created entirely through algorithms, producing hyper-realistic content unlinked to real individuals. In contrast, deepfake pornography alters existing footage of real individuals, often without consent, by superimposing faces or modifying scenes. Hany Farid, a digital image analysis expert, has emphasized these distinctions.

Most deepfake pornography is made using the faces of people who did not consent to their image being used in such a sexual way. In 2023, Sensity, an identity verification company, has found that "96% of deepfakes are sexually explicit and feature women who didn't consent to the creation of the content". In the United States, the TAKE IT DOWN Act was signed into law in May 2025, which addressed non-consensual intimate images as well as deepfake pornography. In the United States, despite 38 states having laws regarding AI CSAM in 2025, several loopholes in enforcement remain, including the lack of definition of "student on student" AI generated deepfake child pornography. The "Deepfake Accountability Act" was introduced to the United States Congress in 2019 but died in 2020. It aimed to make the production and distribution of digitally altered visual media that was not disclosed to be such, a criminal offense. The title specifies that making any sexual, non-consensual altered media with the intent of humiliating or otherwise harming the participants, may be fined, imprisoned for up to 5 years or both. A newer version of bill was introduced in 2021 which would have required any "advanced technological false personation records" to contain a watermark and an audiovisual disclosure to identify and explain any altered audio and visual elements. The bill also includes that failure to disclose this information with intent to harass or humiliate a person with an "advanced technological false personation record" containing sexual content "shall be fined under this title, imprisoned for not more than 5 years, or both." However this bill died in 2023. As of 2023, there was a lack of legislation that specifically addressed deepfake pornography. Instead, the harm caused by its creation and distribution was being addressed by the courts through existing criminal and civil laws. On September 14, 2026, New York District Attorney Alvin Bragg announced that his office had seized twelve websites hosting nonconsensual AI-generated intimate imagery, the largest such seizure to date. Sexually explicit deepfakes had been criminalized in New York since 2023.

This guide takes you through a forensic video analysis workflow covering visual artifacts, metadata, AI-based detection, geometric consistency and spectral analysis. It helps you cross-check findings and reach conclusions that are explainable, reproducible and forensically defensible. Determining whether a video is authentic or manipulated is one of the most pressing challenges in modern forensic media analysis. Public discussion often oversimplifies this issue to a binary question: “Is this a deepfake?”. In reality, true forensic analysis requires a much deeper investigation.

Which anomalies should we look for, and where do we actually begin when a video’s integrity is called into question? 2. System Architecture & Core Working Principles In this post, we outline a scientific workflow for forensically sound video analysis. This methodology is designed not just to detect manipulation, but to generate evidence through an approach that is explainable, reproducible, and legally defensible.

Prerequisite : This guide is written for forensic practitioners who are already familiar with the fundamentals of digital imaging and core forensic methodologies. If you need a refresher, read our overview of the fundamental principles and challenges of working with image and video evidence. Use this workflow to move from an initial suspicion to a documented, evidence-based assessment.

Review the video frame by frame and record exactly where flickering, warping, unstable text or other temporal inconsistencies appear. Examine the container, encoding properties and metadata. Separate meaningful indicators, such as an AIGC (Artificial Intelligence Generated Content) tag, from fields that may simply be missing or poorly populated.

Run AI-based analysis across the video timeline and identify the frames or segments that require closer examination. Treat the output as a lead, not as the conclusion. Check whether shadows, reflections, perspective and object geometry remain coherent within each frame and over time. Examine spectral information in the video or audio to uncover patterns that may not be apparent during normal playback.

Compare the results from each analysis domain. Your conclusion is stronger when several independent traces point in the same direction. The most intuitive place to start is visual analysis. Generative AI models are notorious for introducing physical anomalies, particularly when rendering complex details like human hands, facial features, text, and fine patterns.

In a static image, an artifact is either present or absent. In video, however, anomalies often materialize only briefly, as frame-by-frame inconsistencies, often appearing as a sudden flickering effect. In the example below, look closely at how the subject’s thumbnail warps awkwardly for just a single frame. We can also expose these anomalies by analyzing the video’s spatial continuity. For instance, in this next example, notice how the “STOP” sign suddenly shifts its appearance and orientation between two consecutive frames.

While visual inspection is a vital starting point, it is rarely straightforward in practice. In real-world forensic cases, source videos are often heavily compressed. Distinguishing between a synthetic “AI-generation artifact” and a standard, expected “compression artifact” can be incredibly difficult, even for experienced examiners. Moreover, as generative models continue to rapidly improve, these visual cues will only become scarcer and harder to spot.

The video stream itself is only one piece of the puzzle. It resides within a digital container file packed with specific encoding properties and metadata. Naturally, the next step is to scrutinize the file structure and metadata for any hidden AI footprints. Even when a video’s visual content appears flawless, the underlying metadata can tell a different story. However, interpreting this data requires caution, as certain anomalies can easily mislead an examiner. Take, for example, the metadata extracted from the “broken smartphone” video below: While these fields should ideally contain valid timestamps, seeing null or “zeroed-out” values is not automatically proof of tampering. Metadata population varies wildly across different device brands, operating systems, and camera models.

Fortunately, some generative tools actively leave clear indicators in the file structure, if you know where to look. In the very same “broken smartphone” file, for instance, we can locate an explicit AIGC (Artificial Intelligence Generated Content) tag embedded in the metadata. Note that while finding an AIGC tag is highly significant, it rarely explains the complete history of the media on its own. The tag itself is compatible with several distinct scenarios: the video could be entirely AI-generated, partially manipulated in a specific region, or simply processed or exported using an AI-assisted editing tool.

Another critical piece of the puzzle is using AI to detect AI. Just like we do in our image-based workflows, we can leverage deep learning models to determine if a video’s pixel statistics are consistent with synthetic generation. By applying state-of-the-art image deepfake detection algorithms frame-by-frame across a video, we can isolate exactly which parts of a timeline are suspicious. In the example below, we ran a questioned video through the deepfake detection filters in Amped Authenticate’s Video Mode . The analysis revealed a highly consistent compatibility score (averaging over 80%) with known AI generation models across the flagged frames.

While this automated screening is incredibly powerful, we must remember that deep learning tools are not a silver bullet for forensic investigations. On their own, “black box” neural network outputs struggle to meet strict legal and forensic standards. Rather than serving as final proof, AI-based detection is best used as a triage tool to flag anomalies and guide deeper into more explainable analyses.

This is where checking a video’s physical properties becomes invaluable. Over the last few years, forensic research has shown that AI-generation models may struggle with physical consistency by introducing geometric discrepancies in shadows, perspective lines, and reflections 1 . While these subtle physical inconsistencies can easily trick the human eye, they can be mathematically analyzed and verified using geometric forensic tools (learn more in our blog post on shadow and reflection analysis ). Furthermore, because we are dealing with video, maintaining this physical consistency over time across multiple frames is an incredibly complex task for generative algorithms.

3. Software Implementation, Frameworks & Protocols A classic, pre-AI example perfectly illustrates this vulnerability. Years ago, a viral video showing a golden eagle snatching a child in a park fooled millions of viewers and was even featured on major networks like ABC News . While the visual quality easily deceived the public, a geometric analysis of the shadows cast by the eagle’s wings quickly exposed the hoax. The angle of the light source required to produce the bird’s shadow simply did not match the lighting of the rest of the scene, mathematically proving the video was not authentic. If you are curious about how it works, you can find a more technical description of shadow analysis in this blog post on the subject.

It is worth noting that this viral eagle video could never have been caught by a “deepfake detector” because it was created using older, manual CGI compositing techniques rather than modern generative AI. This distinction is crucial since, when analyzing digital media, we should avoid the dangerous logical fallacy: “If a deepfake detector doesn’t flag it, the video is authentic.” This is why traditional forensic tools, such as geometric and physical analysis, offer a much broader range of applications than deep learning classifiers. The trade-off is that these tools cannot tell you how an inconsistency was made. They cannot distinguish between an image created through highly skilled manual editing and one generated entirely by AI.

Fortunately, in the vast majority of forensic cases, the primary objective is to evaluate the video’s trustworthiness. Pinpointing the exact underlying technology used to alter the scene is often secondary to simply proving that an alteration took place. As a final step in a comprehensive multi-domain video analysis, we should not forget to look beyond what is visible to the naked eye. Academic research has repeatedly highlighted that even visually flawless AI-generated and manipulated content often exposes distinct anomalies in its spectral representation 2 .

A perfect real-world example of this occurred during an episode of the Amped Podcast , where the hosts featured a deepfake voice clone of Arnold Schwarzenegger speaking with Amped Software’s Forensic Director. If you analyze the audio file using a Fourier transform and look at the spectrogram amplitudes, the deception becomes visible as an unexpected frequency peak spike during “Arnold’s” synthesized speech segments.

The voice clone was, of course, created as a joke. However, it also provided a practical demonstration of how spectral analysis can reveal signs of synthetic manipulation. Watch the full Amped Podcast episode to see the example in context. The described workflow is built on a fundamental principle of forensic analysis: while it is relatively easy for a manipulator to conceal a single trace of alteration, producing a fake video that remains perfectly consistent across every aspect is far more difficult.

This is why the proposed methodology focuses on gathering and cross-referencing clues across multiple domains, including visual artifacts, container structure, metadata, AI-based detectors, geometric consistency, and spectral analysis. Combining evidence from these different domains supports a more solid and forensically defensible conclusion. Ultimately, this workflow highlights that answering the seemingly simple question, “Is this video a deepfake?” requires far more than a single click on a deepfake detector. It demands rigorous investigative skills, an objective scientific methodology, and a deep, fundamental understanding of image encoding, compression, and physics.

You can rarely rely on one clue or tool. A sound assessment combines visual inspection, metadata review, AI-based detection, geometric analysis and spectral analysis. The conclusion becomes stronger when findings from different areas support one another. Look for brief frame-to-frame inconsistencies, including flickering, warped facial features, unstable hands, distorted text and changes in fine details. These anomalies may appear for only one or two frames, so normal playback may not reveal them.

Metadata can provide valuable evidence, but it rarely proves the full history of a file on its own. An AIGC tag may indicate AI-generated or AI-assisted content, while missing timestamps or unusual fields may have harmless explanations. Deepfake detectors are useful for screening footage and locating suspicious frames or segments. Their output should guide further examination rather than serve as the conclusion, especially when the model’s decision cannot be fully explained.

No. A video may have been altered through compositing, CGI, editing or other techniques that do not involve generative AI. A detector that finds no deepfake traces cannot rule out other forms of manipulation. Geometric analysis tests whether shadows, reflections, perspective and lighting remain physically consistent. These relationships can be measured rather than judged by eye only, making them useful when visual artifacts are subtle or absent.

Spectral analysis can expose patterns that are not visible during playback. In audio, for example, a spectrogram may reveal unusual frequency behaviour associated with synthetic or manipulated speech. Reading time: 8 min This guide takes you through a forensic video analysis workflow covering visual artifacts, metadata, AI-based detection, geometric consistency and spectral analysis. It helps you cross-check findings and reach conclusions that are explainable, reproducible and forensically defensible. Determining whether a video is authentic or manipulated is one of the most pressing challenges in modern forensic media […] Reading time: 8 min Analyzing suspected deepfakes requires more than detection alone. This post explains a forensically sound workflow for image analysis that helps investigators produce findings that are explainable, reproducible, and defensible.

Reading time: 12 min OSINT investigations are transforming intelligence gathering, from criminal cases to human rights inquiries. Learn how Amped FIVE and Authenticate enhance the accuracy and reliability of digital intelligence. They provide advanced tools for metadata analysis, geolocation, and deepfake detection. Reading time: 5 min The rise of AI-generated CSAM is blurring reality and fueling exploitation. With deepfakes becoming more sophisticated, identifying and combating this abuse is crucial. Discover how forensic tools like Amped Authenticate are helping investigators verify content, fight back, and protect victims.

Forensic video analysis is the scientific examination of video recordings to determine whether they are authentic, manipulated, or entirely generated by artificial intelligence. As synthetic video grows more convincing, forensic methods that analyze frame-level signals, temporal patterns, and compression artifacts have become essential tools for journalists, legal professionals, and platform integrity teams.

4. Practical Deployment, Analysis & Case Studies The demand for reliable video authentication has accelerated sharply. A 2024 report from Sumsub found that deepfake-related fraud attempts increased by 245% year over year, with video deepfakes used in everything from financial scams to political disinformation. The FaceForensics++ benchmark, one of the most widely cited datasets in the field, covers over 1.8 million manipulated images and videos, and the best detection methods now achieve accuracy rates above 95% on controlled data. Real-world conditions, however, remain far more challenging.

This guide covers the core methods used in forensic video analysis, explains how each technique works at a technical level, and shows where the field is headed as AI-generated video continues to evolve. Forensic video analysis sits at the intersection of signal processing, computer vision, and digital forensics. The goal is straightforward: given a video file, determine what happened to it. Was it captured by a real camera? Has it been edited? Were faces swapped? Was any portion generated by an AI model? The methods fall into two broad categories. Active methods rely on signals that were embedded during creation, such as digital watermarks or C2PA content credentials. Passive methods rely entirely on signals present in the video itself, requiring no cooperation from the creator. AFIP's forensic approach emphasizes passive methods because they work regardless of whether the creator chose to label their content.

Image forensics analyzes a single frame. Video forensics analyzes thousands of frames and the relationships between them. This adds a temporal dimension that both complicates and strengthens the analysis. A single deepfake frame might look perfect. But play 30 of those frames per second, and inconsistencies emerge: lighting flickers where it should be stable, face boundaries shift unpredictably, blinking patterns fail to match normal human physiology. The temporal dimension is a forensic advantage because maintaining perfect consistency across time is significantly harder than generating a single convincing frame.

Video also carries audio, and audio-visual synchronization is another layer of analysis entirely. A face-swapped video might have flawless visual quality but fail when the lip movements are compared against the speech waveform at the phoneme level. For forensic video analysis to hold up in legal proceedings, the chain of custody must be preserved from the moment the video is acquired. This means documenting who handled the file, what tools were used, and what hash values were computed at each stage. ISO 27037 provides the international standard for digital evidence handling, and SWGDE (the Scientific Working Group on Digital Evidence) publishes specific best practices for multimedia evidence.

Courts have accepted forensic video analysis in proceedings ranging from intellectual property disputes to criminal cases involving surveillance footage. The key requirement is reproducibility: another qualified examiner, given the same file and the same tools, should reach the same conclusions. Forensic video analysis draws on multiple detection methods, each targeting a different type of artifact. No single method catches everything. Effective analysis combines several techniques and weighs the results together.

Every legitimate video exhibits certain frame-to-frame consistencies that are difficult to fake. Lighting conditions change gradually across frames. Noise patterns from the camera sensor repeat with statistical regularity. Color balance shifts follow predictable patterns tied to the camera's automatic white balance algorithm. Frame-level consistency analysis examines these patterns by comparing statistical properties across consecutive frames and flagging sudden discontinuities. A spliced video, where a segment from one recording is inserted into another, will often show abrupt changes in noise floor, color temperature, or quantization patterns at the edit points.

The analysis typically works on small blocks (8x8 or 16x16 pixel patches) rather than full frames, because manipulation often affects only a portion of the image. This block-level approach can localize the tampered region rather than simply flagging the entire frame. Temporal coherence refers to the smooth, physically consistent progression of visual elements over time. Real-world physics imposes constraints that AI generators struggle to maintain perfectly. Shadows move in predictable arcs as light sources shift.

Reflections in eyes, windows, and other surfaces maintain geometric relationships with the objects they mirror. Hair, clothing, and other soft materials follow fluid dynamics that current generation models approximate but do not perfectly simulate. Forensic temporal analysis builds a model of expected motion and appearance across frames, then measures deviations from that model. The most effective implementations use optical flow estimation to track the movement of every pixel region across time. When a face has been swapped, the optical flow field around the face boundary will often show discontinuities that are invisible to the human eye but statistically significant when measured.

Modern video codecs like H.264, H.265, and VP9 compress video by dividing frames into blocks and predicting each block from previously decoded data. This process leaves characteristic artifacts: blocking patterns at consistent boundaries, quantization noise that follows specific distributions, and prediction residuals that carry the codec's mathematical fingerprint. When a video is edited and re-encoded, these artifacts get layered. The original encoding artifacts interact with the new encoding in ways that produce detectable double-compression signatures. A frame that was decoded, modified, and re-encoded will have a different quantization pattern than a frame that was encoded only once.

H.264 and H.265 use a hybrid coding model that combines intra-frame prediction with inter-frame prediction. Forensic analysis targets the quantization parameter (QP) values and DCT coefficient distributions in each macroblock. Double compression shifts these distributions in measurable ways, particularly when the two compression passes used different QP values or different block boundaries.

This method is especially useful for detecting videos where only a portion of the frames have been modified. Unaltered frames will show single-compression statistics, while tampered frames will show the telltale double-compression signature. Video codecs encode motion by computing motion vectors: directional offsets that describe how each block in the current frame relates to the reference frame. These motion vectors follow physical constraints. A camera pan produces a uniform field of vectors pointing in the same direction. A person walking produces vectors that follow the biomechanics of human gait.

AI-generated video often produces motion that looks correct at the surface level but contains subtle anomalies in the motion vector field. The vectors might be slightly too smooth (lacking the natural noise of real camera motion), or they might show inconsistencies where different objects in the scene move in ways that violate physical proximity constraints. AI-generated video presents a distinct forensic challenge because it may not have been captured by a camera at all. There is no sensor noise pattern to analyze, no lens distortion to model, and no original codec fingerprint. Instead, forensic analysis of AI video targets the artifacts left by the generation process itself.

Generative adversarial networks leave spectral fingerprints that persist even after compression and resizing. In the frequency domain, GAN-generated frames show characteristic peaks at specific frequencies that natural images do not exhibit. These peaks arise from the upsampling operations (transposed convolutions) used in the generator network. 5. Performance Evaluation & Extended Technical Observations For video, these spectral fingerprints should be consistent across frames if the entire video was generated by the same model. Forensic analysis checks both the presence of these fingerprints and their consistency over time. A mixed video, where some frames are real and others are generated, will show spectral signature shifts at the transition points.

Face-swapping, the most common form of video deepfake, replaces the face of one person with another while keeping the body, background, and audio intact. This process creates several categories of artifacts: When speech audio is paired with a face-swapped or AI-generated face, the lip movements must match the phonemes being spoken. This is a hard problem. Certain phonemes, like bilabial plosives (/p/, /b/, /m/) require the lips to close completely.

Others, like labiodental fricatives (/f/, /v/) require the lower lip to contact the upper teeth. Forensic lip-sync analysis uses phoneme-level alignment to measure whether the visual mouth shapes match the expected articulation. The analysis computes a frame-by-frame correspondence score. Real video typically scores above 0.85 on normalized alignment metrics, while deepfakes often fall between 0.55 and 0.75 due to imperfect mouth modeling.

Humans blink approximately 15 to 20 times per minute, with each blink lasting 100 to 150 milliseconds. Early deepfake generation models, particularly those from 2017 to 2019, produced faces that rarely blinked at all, making blink-rate analysis an effective detector. Modern models have largely corrected this, but blink analysis still catches lower-quality deepfakes. A more robust biological signal is remote photoplethysmography (rPPG), which detects the subtle color changes in facial skin caused by blood flow. Real faces show a periodic pulse signal that matches typical heart rates (60 to 100 bpm). AI-generated faces typically lack this signal entirely, or produce a signal with incorrect frequency characteristics.

Tracks eyelid closure frequency, duration, and bilateral symmetry. Effective against older models, less reliable against state-of-the-art generators. Measures periodic color fluctuations from blood flow. Absent in most AI-generated faces. Robust even after moderate compression.

Analyzes eye movement patterns for saccades, fixation, and pupil dilation. AI faces often show unrealistic gaze stability or synchronization errors between eyes. Detects chest and shoulder movement associated with respiration. Frequency and regularity mismatches indicate potential manipulation of the body region.

No single detection method catches everything. The forensic advantage comes from combining multiple independent signals into an evidence-weighted verdict. The distinction between passive and active detection is fundamental to understanding the current landscape of video authentication. Each approach has strengths and meaningful limitations.

Every video file carries metadata: container-level information (format, duration, creation date), codec parameters (profile, level, bitrate), and sometimes device-specific data (camera model, GPS coordinates, firmware version). Forensic analysis examines this metadata for inconsistencies. A video that claims to be shot on an iPhone 15 Pro but uses codec settings that no iPhone produces is immediately suspicious. A file with a creation timestamp that predates the events depicted requires explanation. Metadata analysis is fast and can provide early triage, but it is easily defeated by stripping or forging metadata, which is trivial to do with common video editing tools.

Passive forensic analysis examines the pixel data itself, independent of any metadata or embedded signals. This approach works on any video, regardless of its origin or how many times it has been re-encoded, screenshotted, or re-uploaded. The methods described in the previous sections, including frame consistency, temporal coherence, compression analysis, and biological signal detection, are all passive techniques.

The strength of passive analysis is that it cannot be opted out of. A manipulated video cannot choose to remove the artifacts that forensic analysis detects, because those artifacts are inherent to the manipulation process itself. The limitation is that passive methods face an ongoing arms race: as generators improve, the artifacts become subtler, and detection methods must evolve to match.

Active methods, such as digital watermarks and C2PA content credentials, embed information into the video at the point of creation. C2PA, backed by a coalition including Adobe, Google, and Microsoft, attaches a cryptographic manifest to the file that records how and where the content was created. Active methods provide strong provenance when they are present. The problem is that they require voluntary participation.

Most social media platforms strip metadata and C2PA manifests during upload. And a bad actor creating manipulated video has no incentive to embed watermarks or sign their content with credentials. This is why AFIP advocates a complementary model: use active methods where available, but always have passive forensic analysis as the fallback. C2PA tells you what the creator claims. Forensic analysis tells you what the evidence shows. Both are valuable, but only forensic analysis works when the creator does not cooperate or when metadata has been stripped during sharing.

Forensic video analysis requires both specialized software and a structured workflow. The field includes open-source research tools, commercial platforms, and integrated solutions like AFIP's forensic analysis pipeline. The research community has produced several open-source tools that implement specific detection methods. FaceForensics++ provides a benchmark dataset and baseline detectors for face manipulation. MediaPipe from Google offers face mesh extraction that can be used for geometric analysis. FFmpeg , while not a forensic tool per se, provides the foundation for frame extraction, codec analysis, and metadata examination that underpins most forensic workflows.

These tools are powerful but require technical expertise to use effectively. They are best suited for research environments and organizations with dedicated forensic analysts. AFIP's video forensic analysis combines multiple detection methods into a single pipeline that produces an evidence-weighted confidence score. Rather than relying on any single detector, the pipeline runs frame consistency, temporal coherence, compression forensics, biological signal analysis, and audio-visual synchronization checks in parallel, then fuses the results using a Bayesian evidence weighting model.

Harvested Online Documentation & Reference Sources • https://en.wikipedia.org/wiki/Audio%20deepfake • https://en.wikipedia.org/wiki/Deepfake • https://en.wikipedia.org/wiki/Deepfake%20pornography • https://blog.ampedsoftware.com/2026/07/28/deepfake-forensics-workflow-for-video-analysis • https://afip.org/research/video-forensics/