How Are AI Deepfakes Created: A Complete Technical Guide to Synthetic Media, From GANs to Real Time Voice Cloning

Synthetic media fraud cost United States victims over $1.1 billion in 2025, according to Forbes' analysis of federal fraud data, nearly tripling the $360 million lost the year prior. That growth curve makes understanding how are AI deepfakes created a practical requirement for any organization building defenses against AI-powered social engineering, not an academic exercise reserved for machine learning researchers.

This guide traces the complete deepfake creation pipeline from raw data collection through the neural network architectures that power generation, including GANs, autoencoders, diffusion models, and transformer-based approaches. It also covers voice cloning, computing requirements, post-processing techniques, and the detection methods built to counter them.

This guide covers:

• How AI deepfakes are created, step by step, from data collection through post-processing
• The neural network architectures, including GANs, autoencoders, and diffusion models, that power modern synthetic media
• How voice cloning and real-time deepfake video threaten executive impersonation and business email compromise
• The legal frameworks, detection tools, and cybersecurity awareness training strategies organizations use to build resilience

Security teams that do not understand the deepfake creation pipeline cannot train employees to spot its output. Adaptive Security's phishing simulations build detection instincts before employees face the real thing.

Take a self-guided tour

What Is an AI Deepfake?

An AI deepfake is a synthetically generated or manipulated piece of media, video, audio, or imagery, created using artificial intelligence to realistically depict a person saying or doing something that never happened. Neural networks trained on real recordings produce these fabricated outputs, which now approach a level of realism that defeats both casual human judgment and biometric verification systems. Understanding how are AI deepfakes created starts with defining what separates a deepfake from an ordinary edited photo or video.

The Core Definition and What Qualifies as a Deepfake

Not every synthetic or altered piece of media qualifies as a deepfake. The term combines "deep learning" and "fake," and its defining characteristic is the use of deep neural networks, specifically architectures like generative adversarial networks (GANs), autoencoders, and diffusion models, to generate or substantially alter media. A deepfake is not a video with a misleading caption or an image run through a beautification filter.

Three criteria separate a deepfake from other forms of manipulated media. First, the alteration must be AI-driven: machine learning models trained on source data, not manual editing in timeline-based software. Second, the output must be realistically deceptive, meaning a reasonable observer would struggle to distinguish it from genuine media.

Third, the subject is usually a specific, identifiable person rather than a fictional character.

This last point makes deepfakes uniquely dangerous as a social engineering tool. A cyberattacker is not inventing a persona but hijacking one the target already trusts. A finance employee does not need to be convinced that a stranger needs a wire transfer; the employee needs to believe a familiar executive asked for one.

According to a 2024 Regula survey on identity verification, cited in a Statista chart analysis, 46% of fraud experts had already encountered synthetic identity fraud, 37% had faced voice deepfakes, and 29% had dealt with video deepfakes. These figures underscore how quickly deepfake attack vectors have penetrated the fraud landscape across industries.

Deepfakes vs. Shallowfakes and Other Synthetic Media

Not all manipulated media is a deepfake, and the distinction shapes both detection strategy and organizational defense. Shallowfakes, sometimes called cheapfakes, are media altered through low-effort, non-AI methods: a video slowed to make a speaker appear intoxicated, a clip taken out of chronological context, or audio dubbed over existing footage. These manipulations require no machine learning and can be executed with basic editing tools.

"Deepfakes use artificial intelligence to seamlessly alter faces, mimic voices, or fabricate actions in videos," said Nirupam Roy, Assistant Professor of Computer Science at the University of Maryland and a core faculty member at the Maryland Cybersecurity Center. "Shallowfakes rely less on complex editing techniques and more on connecting partial truths to small lies."

Deepfakes also occupy a specific subset within the broader category of synthetic media, which encompasses any AI-generated content, including fictional characters and computer-generated imagery for entertainment. What separates a deepfake from these other forms is impersonation. An AI-generated landscape is synthetic media but not a deepfake.

A synthetically generated video of a real chief financial officer authorizing a fraudulent transaction is a deepfake, and that impersonation vector is what makes deepfakes a cybersecurity problem rather than a content authenticity concern.

For security teams, the practical difference shapes response strategy. Shallowfakes can often be debunked through reverse image searches or source verification, while deepfakes generated by diffusion models frequently leave no detectable metadata trail. This gap means organizations cannot rely on detection alone.

Teams must train employees to confirm high-stakes instructions through a second trusted channel regardless of how convincing the initial communication appears, a skill built through phishing simulations that rehearse deepfake video and voice scenarios before a cyberattacker tests it for real.

The Evolution of Deepfake Technology From 2014 to the Present

The theoretical foundation for deepfakes was laid in June 2014, when Ian Goodfellow and colleagues at the Université de Montréal published their paper introducing generative adversarial networks. The GAN architecture pits two neural networks, a generator and a discriminator, against each other in an adversarial game.

The generator produces synthetic content, and the discriminator attempts to distinguish it from real data. Through iterative competition, both networks improve until the generator produces outputs the discriminator can no longer reliably classify.

Goodfellow's framework did not mention deepfakes by name; a Reddit user coined the term in 2017. But the paper provided the algorithmic engine that would power them. The face-swap era that followed, roughly 2017 through 2019, was characterized by autoencoder-based tools requiring substantial technical expertise and large volumes of source imagery.

Outputs were often visibly flawed, with mismatched lighting and inconsistent blinking as common giveaways. Only technically proficient cyberattackers with significant resources could target high-value victims at that stage.

Between 2020 and 2023, GAN-based generation improved steadily. Voice cloning tools reduced audio sample requirements from hours of speech to under a minute. The quantum leap arrived in 2024 with the commercialization of diffusion models, the same architecture behind image generators, now applied to video. Diffusion models work by gradually denoising random static into structured imagery rather than competing adversarially, a process that produces more temporally stable and visually convincing output.

By 2025, open-source diffusion-based video models became available to anyone with a consumer-grade GPU. Voice cloning required only a few seconds of source audio. Real-time facial reenactment tools could animate a static photo during a live meeting.

The $25.6 million Arup wire fraud in early 2024, in which an employee joined a video conference where every other participant was a deepfake, was executed with commercially available tools rather than classified technology. That single incident crystallized that deepfakes had become a present-day operational risk every organization must account for in its human-layer defense strategy.

Defeating executive verification with a deepfake no longer requires specialized skill or expensive equipment. Adaptive Security's awareness training teaches employees to recognize manipulation before it reaches a wire transfer approval.

Explore the platform

The Step-by-Step Deepfake Creation Process

Creating a convincing AI deepfake moves through five sequential stages: gathering images or audio of the target, preprocessing that data into a consistent training format, training a neural network to learn the target's appearance or voice, generating synthetic output frame by frame, and blending the result into source media to erase boundary artifacts. Each stage presents its own failure points; a sparse dataset produces a blurry clone, poor alignment creates visual glitches, and skipped post-processing leaves a detectable seam between synthetic and real elements. Understanding how AI deepfakes are created at this granular level is what allows security teams to identify where the process is most vulnerable to disruption.

1. Data Collection and Source Material Gathering

Every deepfake begins with reconnaissance. The cyberattacker assembles a dataset of the target's face or voice, pulling from LinkedIn photos, YouTube conference talks, podcast appearances, and earnings call recordings.

For voice cloning, a few seconds of clean reference audio can produce a usable clone. For face-swap video, a single high-resolution reference photo is often sufficient, though higher-fidelity outputs benefit from larger datasets spanning varied angles and lighting.

This stage is why open-source intelligence hygiene has become central to the deepfake cyber threat. A 2024 study published in Human-Intelligent Systems Integration described how deepfake creation relies on neural networks that analyze extensive datasets to learn facial features, expressions, and voice patterns. Publicly available content, especially for executives who speak at conferences or maintain active social media presences, provides exactly the type of extensive dataset these networks need.

The most common failure at this stage is dataset sparsity. If training images all show the subject from the same angle or lighting, the resulting deepfake collapses when rendering the face under different conditions.

Cyberattackers mitigate this by scraping content across platforms and time. Limiting an executive's publicly available video and audio footprint reduces the raw material available for a convincing deepfake.

2. Preprocessing and Face Alignment

Raw collected images are not ready for training. Preprocessing converts this chaotic dataset into a consistent, machine-readable format through a sequence of automated steps. Face detection locates the target face in each frame.

Landmark extraction identifies key facial features such as the eyes, nose, and mouth corners. Alignment then rotates and scales each face so those landmarks occupy the same position across every image.

Cropping isolates the face region and discards irrelevant background. Normalization adjusts brightness and color balance so a dimly lit conference photo and a bright outdoor selfie produce comparable pixel values. A 2025 survey in the Journal of Imaging found that preprocessing in deepfake creation includes face detection, alignment, and normalization to ensure data is suitable for training.

Skipping or rushing this phase produces models with distorted faces, eyes that fail to track with head angle, or skin tones that shift between frames. The quality ceiling for any deepfake is set at this stage; no amount of model training can compensate for poorly aligned input data.

3. Model Training and the Generation Loop

With the preprocessed dataset in hand, the cyberattacker feeds it into a neural network architecture, typically an autoencoder, a GAN, or increasingly, a diffusion model. The choice of architecture shapes everything downstream: training time, output quality, and the type of artifacts a detector might later identify.

Autoencoders compress face images into a compact latent representation and learn to reconstruct one specific person's face from that representation. In a face swap setup, two autoencoders share a single encoder. One decoder learns to reconstruct Person A's face; the other learns Person B's.

At inference, the cyberattacker passes Person B's video through the shared encoder to capture pose and expression, then routes the latent code through Person A's decoder to render Person A in Person B's position.

Training is computationally intensive but no longer a barrier. A pre-trained foundation model can be conditioned on a handful of reference images at inference time, collapsing what once required a GPU cluster and days of training into minutes on a single consumer GPU.

4. Post-Processing and Refinement

Raw generated output is rarely usable as-is. The synthetic face must be composited into the source video, which introduces visible seams at the jawline, hairline, and ears. Poisson image editing and similar blending algorithms smooth these transitions by matching pixel intensity gradients across the boundary.

Color grading then matches the synthetic face to the lighting of the original scene, since a face generated under studio lighting looks conspicuously wrong when dropped into warm office lighting. Cyberattackers adjust hue, saturation, and luminance curves frame by frame to achieve consistency, and frame interpolation smooths temporal inconsistencies that betray synthetic output.

"Tools once reserved for governments or well-funded organizations are now freely available, radically expanding the threat landscape," said Dr. Hany Farid, professor at the University of California, Berkeley's School of Information and a longtime authority on digital media forensics. "The foundation models are good enough out of the box." Defenders who understand where deepfakes fail, in sparse data, sloppy alignment, and visible blending artifacts, are better equipped to train employees to spot them.

Employees who have never seen how a deepfake seam actually looks cannot flag one in a live video call. Adaptive Security's phishing simulations use AI-generated deepfake video and voice to build practiced detection.

Take a self-guided tour

The AI Architectures That Power Deepfake Creation

Four distinct neural network architectures drive the deepfake ecosystem, each solving a different part of the synthetic media puzzle. A 2025 comprehensive survey in the Journal of Imaging documented that generative adversarial networks and diffusion models now represent the two dominant paradigms for producing photorealistic deepfake imagery, while autoencoder-based face-swapping and motion-transfer models handle identity replacement and video animation. These architectures differ in how they learn, what they produce, and the compute resources they demand, and cyberattackers increasingly chain them together in multi-stage pipelines that defeat single-point detection.

Generative Adversarial Networks Explained

GANs operate on an adversarial principle: two neural networks compete until the forgeries become indistinguishable from reality. The generator synthesizes fake images from random noise, while the discriminator tries to classify each output as real or synthetic. This arms race continues for millions of training iterations until the discriminator performs no better than a coin flip.

Introduced by Ian Goodfellow and colleagues in 2014, GANs dominated deepfake generation for nearly a decade. StyleGAN and its successors became the gold standard for face synthesis, generating portraits with controllable pose, lighting, and expression.

A 2026 Lancaster University study led by researcher Alexis McGuire, with coauthors from Lancaster, Stanford, and UC Berkeley, found that GAN-generated faces now receive higher trustworthiness ratings than photographs of real people. Participants rated GAN faces 4.36 and diffusion model faces 4.70 on a 1 to 7 trustworthiness scale, compared to 4.03 for real faces.

GANs excel at generating entirely new faces that belong to no real person, useful for fabricating fake social media profiles at scale. Their primary weaknesses include training instability and mode collapse, where the generator produces only a narrow range of outputs. Pre-trained models and fine-tuning techniques have significantly lowered the technical barrier that once limited GAN use to well-resourced threat actors.

Autoencoders, Latent Representations, and Face Swapping

The classic deepfake face swap that launched the synthetic media era on Reddit in 2017 relies on autoencoders rather than GANs. An autoencoder is a neural network architecture built from two connected components.

An encoder compresses an input face into a compact latent representation, a mathematical abstraction capturing essential facial features like structure, pose, and expression. A decoder then reconstructs a face from that compressed code.

The face-swapping trick works through shared-encoder training, where two decoders are trained on separate identities but share the same encoder. The shared encoder learns a universal facial representation that abstracts away identity-specific details while preserving pose and expression.

Autoencoders produce convincing face swaps with moderate compute requirements; a consumer GPU and a few thousand face images are sufficient. Their weakness is output consistency across extreme pose or lighting variation, which often produces visible blending artifacts around the jawline and hair.

Diffusion Models and the 2024-2025 Quality Revolution

Diffusion models work backward from chaos. Training begins by taking real images and progressively adding noise across many steps until the original is destroyed, leaving pure random noise. The model then learns to reverse this degradation, iteratively denoising toward a coherent image.

Unlike GANs, which generate an image in a single forward pass, diffusion models refine output across many sequential steps, producing far more detailed and photorealistic results.

The quality leap became undeniable in 2024. Diffusion-based models advanced so rapidly that AI-generated faces, voices, and full-body performances became nearly indistinguishable from genuine recordings. Where GAN-generated faces often exhibited telltale artifacts in eye symmetry and background consistency, diffusion models eliminated many of those signatures.

Training remains computationally intensive, and each generation requires multiple inference passes. Diffusion models proved more stable to train than GANs because they optimize a likelihood-based objective rather than an adversarial game, virtually eliminating mode collapse.

First-Order Motion Models, Transformers, and CNNs

Still images alone do not make a deepfake video convincing. Motion is what triggers human trust. Micro-expressions and head turns signal authenticity in ways a static frame cannot. First-order motion models solve this by decoupling appearance from motion entirely, extracting a motion field from a driving video and applying it to a static source image so the target face moves exactly like the driver.

Convolutional neural networks handle facial landmark detection and motion field estimation. Transformer architectures process sequential frames with attention mechanisms that maintain temporal coherence, ensuring facial features persist consistently rather than flickering unpredictably. Together, these architectures create the temporal backbone that transforms a single-frame forgery into a multi-channel cyberattack capable of sustaining a live meeting, as occurred in the Arup deepfake fraud, where every participant on a multi-person video conference was synthetic.

Researchers led by Irene Amerini, in a 2025 deepfake forensics survey, noted that recent diffusion model advances have demonstrated state-of-the-art performance in image synthesis, rivaling and sometimes surpassing GANs in quality and diversity. Defenders who understand these models' distinct fingerprints can build detection strategies that target the specific artifacts each architecture leaves behind, from GANs' convolutional traces to diffusion models' frequency-domain signatures.

Recognizing a GAN artifact versus a diffusion signature does not help a finance employee under pressure on a live call. Adaptive Security turns that architectural knowledge into rehearsed behavior through realistic phishing simulations.

Book a demo

How Voice Cloning and Audio Deepfakes Are Created

Voice cloning starts with harvesting clean audio samples of a target speaker, then feeding that data through a neural speech synthesis pipeline that separates vocal identity from linguistic content and recombines them into entirely new sentences. Modern zero-shot systems need only a brief audio sample to produce a recognizable clone, and high-fidelity replicas emerge from minutes of clean speech. The same technology powers live voice conversion cyberattacks, in which a fraudster's spoken words are re-rendered in the target's voice during a real-time phone call.

1. Voice Data Collection and Preprocessing

Every voice clone begins with source material, and cyberattackers have an abundance of it. Criminals harvest audio from earnings calls, conference keynotes, podcast appearances, and voicemail greetings.

According to McAfee's 2023 Artificial Imposters: Cybercriminals Turn to AI Voice Cloning for a New Breed of Scam report, just three seconds of clean audio can produce an 85% voice match, and 53% of adults share their voice online at least once per week. In practice, nearly every executive with a public presence has already published sufficient training data for a usable clone.

The preprocessing pipeline cleans and standardizes this raw audio before it reaches the neural model. Background noise is stripped using deep learning-based denoising algorithms, audio is resampled to a consistent rate, and amplitude is normalized so the model receives uniform input regardless of source quality. These steps are fully automated in modern tooling and require no specialized audio engineering knowledge.

The quality of source audio directly determines clone fidelity. A brief sample yields a recognizable but noticeably synthetic voice, adequate for a short vishing call where urgency suppresses scrutiny. A 2023 University College London study, Warning: Humans cannot reliably detect speech deepfakes, published in PLOS ONE, found that humans correctly identify deepfake speech only 73% of the time, meaning more than one in four listeners cannot tell a cloned voice from a real one.

That gap between what the ear can catch and what a trained neural network can produce is exactly what makes vishing calls harder to defend against than ordinary email phishing.

Deepfake-enabled vishing cyberattacks surged by over 1,600% in the first quarter of 2025 versus the fourth quarter of 2024, according to Right-Hand AI's research on the state of deepfake vishing, driven by how trivially the data collection step has become. Finance team members authorized to approve wire transfers present the highest-value targets because a cloned voice combined with payment authority converts directly to financial loss.

2. Neural Speech Synthesis and Voice Conversion Techniques

Modern voice cloning runs a three-stage pipeline refined to the point where off-the-shelf tools produce convincing results in seconds. The first stage is the speaker encoder, a neural network that ingests reference audio and outputs a speaker embedding, a numerical representation of the voice's pitch distribution, formant patterns, and tonal range. Because the encoder is trained to recognize whether two clips come from the same person rather than to recognize specific words, it generalizes to voices it has never encountered.

The second stage is the synthesis model, which generates an acoustic representation of target speech by combining the speaker embedding with text converted into phonemes. Architectures such as Tacotron 2 use sequence-to-sequence learning with attention mechanisms to align phonemes to spectrogram frames, while newer models use feed-forward transformers for faster, parallel generation.

The third stage is the vocoder, which converts the spectrogram into an audio waveform. Neural vocoders reconstruct waveforms with near-lossless fidelity, which is why modern clones sound smooth rather than robotic. WaveNet pioneered this approach, but modern feed-forward vocoders achieve comparable quality with dramatically lower latency, enabling real-time synthesis.

Voice conversion takes a different path and poses the more immediate cyber threat for live cyberattacks. Instead of generating speech from text, it takes one speaker's live recording and re-renders it in the target voice, preserving original words and pacing. This latency-optimized approach enables the real-time executive impersonation cyberattacks that have become the dominant deepfake-enabled fraud vector, where a fraudster's words are output live in a CFO's voice during an active phone call.

3. NLP, Lip-Syncing, and the Pursuit of Realism

Raw voice cloning produces accurate timbre and pitch, but without natural language processing, the result sounds flat. NLP models govern the prosodic layer: intonation, stress patterns, and emotional coloring. Without these cues, a synthetic voice delivers words without meaning; with them, it can sound urgent or authoritative, precisely the emotional registers that social engineering exploits.

Modern text-to-speech systems integrate prosody prediction directly into the synthesis pipeline, learning mappings between text semantics and acoustic prosody from a large corpora of expressive speech. Cyberattackers exploit this by scripting vishing calls with high-urgency language calibrated to suppress verification instincts, amplified by vocal authenticity that traditional social engineering could not achieve.

When audio deepfakes combine with video, lip-syncing technology aligns the generated speech waveform with mouth movements. Recurrent neural networks learn the mapping between phoneme sequences and viseme sequences, the visual equivalent of phonemes representing distinct mouth shapes. More recent transformer-based sequence models have improved synchronization accuracy, reducing the misalignment that once made deepfake video easy to spot.

The Arup fraud demonstrated the combined power of these techniques. A finance employee joined a video call where every participant, including the chief financial officer, was a synthesized deepfake with matched audio and video, and the employee authorized 15 wire transfers totaling $25.6 million, roughly HK$200 million, before discovering the deception. The cyberattack succeeded because every channel of verification confirmed the deception simultaneously.

A cloned voice paired with an urgent tone defeats verification instincts organizations spent decades building. Adaptive Security's multi-channel phishing simulations rehearse voice and video deepfake scenarios until verification becomes reflex.

Take a self-guided tour

What Makes Deepfakes Convincing: Data, Compute, and Post-Processing

A deepfake's believability depends on three resource dimensions: the quantity and quality of source data, the computing power available to train or run the model, and the post-processing techniques applied to clean up visible artifacts. The gap between a crude face swap that anyone spots and a photorealistic impersonation that fools experienced professionals is rarely about the underlying AI model; it is almost always about how these three resources are deployed. Publicly available conference footage and social media content supplied enough data to build convincing replicas of multiple Arup executives, and the rest was compute and polish.

Minimum Viable Data Requirements by Quality Tier

The amount of source material needed to produce a deepfake has collapsed. What once required hours of studio footage now works from a handful of images or seconds of audio. Modern zero-shot systems extract identity embeddings from pre-trained models rather than training a dedicated model per target, though quality still tracks closely with how much data is available.

• Low-quality tier: A single clear photograph is sufficient. Cybersecurity researchers have warned that as few as 20 publicly available images can produce a convincing deepfake of a specific person, a threshold that virtually every professional with a LinkedIn profile has already exceeded. Output at this tier is recognizable but marred by boundary artifacts and flickering.
• Medium-quality tier: This tier demands hundreds of images with consistent lighting and varied angles. Output achieves smoother facial integration and maintains identity stability across short clips, though close inspection still reveals telltale signs around the hairline and jaw.
• High-quality photorealistic tier: This tier requires thousands of multi-angle images plus hours of video covering different lighting conditions and emotional ranges. The result maintains identity consistency across head turns and lighting shifts, making casual scrutiny unreliable as a detection method.

A 2024 meta-analysis of 56 studies published in Computers in Human Behavior Reports found overall human deepfake detection accuracy averages just 55.54%, barely above chance, and that figure drops sharply against high-quality video specifically, compared to a somewhat higher rate for audio-only speech identification. The critical security insight is that the data floor is not the practical barrier it once was; every executive who has appeared in a recorded earnings call has already published more than enough source material to feed a high-quality pipeline.

Computing Power, Training Time, and Cost

Computing requirements for deepfake generation split into two categories: running pre-trained models for inference, and training custom models from scratch. Consumer-grade inference runs comfortably on a mid-range gaming GPU, performing real-time face swaps with minimal added latency. The model has already been trained on millions of faces during pre-training, and the user is simply supplying a reference image for identity extraction.

Commercial deepfake-as-a-service platforms abstract even the hardware requirement, offering photorealistic talking avatars and voice clones through simple subscription pricing. These services eliminate the technical skill barrier entirely; a threat actor needs no machine learning knowledge, no GPU, and no dataset beyond a few publicly sourced images or audio clips.

Custom model training occupies the opposite end of the spectrum. Training a high-fidelity face-swap model on a large, identity-specific dataset typically costs several thousand dollars at on-demand cloud GPU pricing, though that figure remains well within reach of organized fraud operations targeting high-value enterprises, especially when a single successful cyberattack can yield millions. The asymmetry is stark: a cyberattacker's marginal cost per impersonation approaches zero once a model is trained, while detection infrastructure must operate continuously across every communication channel.

Post-Processing Techniques That Eliminate Visual Artifacts

Raw generation output from even the best models is rarely the final product. The difference between an obviously fake deepfake and one that passes casual scrutiny almost always comes down to the post-processing pipeline applied after synthesis.

Eye-blink correction addresses one of the earliest and most persistent deepfake tells, since early GAN-based models produced faces with unnatural blinking patterns. Modern post-processing pipelines detect and correct blink frequency by interpolating realistic blink frames from a reference library. Skin texture blending restores the fine-grained detail that generation models tend to smooth away, since diffusion models and GANs often produce faces with a telltale plastic quality.

Lighting normalization corrects one of the hardest problems in face replacement: matching illumination conditions between the source face and target scene. Post-processing pipelines perform histogram matching to align color temperature and luminance, then apply gradient-domain blending to smooth the transition. Boundary feathering addresses the most visually jarring artifact, a hard seam where the synthetic face meets the original neck and hair, using Poisson image blending to progressively fade edge pixels into the target frame.

Temporal smoothing eliminates frame-to-frame flicker that betrays per-frame generation, averaging feature values and color histograms across adjacent frames to enforce smooth transitions. Color grading provides the final polish, applying global color temperature and contrast adjustments that unify the synthetic face with the target footage. This step is computationally cheap but disproportionately important, since a face slightly off in color temperature triggers suspicion faster than almost any other artifact.

The cumulative effect is why visual clues that once exposed deepfakes have largely stopped working. Post-processing does not make a deepfake perfect; detection systems analyzing frequency-domain fingerprints can still flag synthetic output. But it reliably pushes quality past the threshold where casual human observation fails, which is why verification protocols must operate on channels independent of visual or auditory trust.

Six specific post-processing techniques now defeat the visual tells finance teams learned to spot a decade ago. Adaptive Security exposes employees to post-processed deepfake content in a controlled drill, not a real cyberattack.

Take a self-guided tour

Tools, Software, and Platforms for Creating Deepfakes

The deepfake creation landscape has split into three distinct tiers that collectively put convincing synthetic media within reach of virtually anyone. Free mobile applications swap faces in seconds, open-source frameworks give technically skilled users frame-by-frame control, and deepfake-as-a-service platforms now sell on-demand voice cloning through simple web interfaces. Dark web posts referencing deepfake-as-a-service reached 924 between January and May 2026, up 39% from the 663 posts recorded during the same period the prior year, according to a 2026 NordStellar analysis of dark web activity.

Consumer-Grade Applications for Instant Deepfake Creation

Consumer face-swap applications dominate the most accessible tier, offering drag-and-drop interfaces that require zero technical expertise and deliver results in under a minute. A user uploads a selfie, selects a target image or video, and the app handles face detection, alignment, and blending automatically. These applications run on GANs pretrained on massive datasets, meaning the heavy computation happened before the app ever reached the user's phone.

The tradeoff is predictability. Consumer apps constrain users to predefined templates, face swaps, aging filters, or style transfers, and cannot produce novel poses or clone a specific voice. For casual misuse or social media deception, however, this tier is more than sufficient.

Advancements in generative AI have accelerated the creation of deepfakes and made them increasingly realistic, lowering the barrier to entry for cybercriminals with no technical background.

Open-Source Frameworks and Research-Grade Tools

Open-source frameworks such as DeepFaceLab and FaceSwap are the most widely used tools for technically skilled creators, each providing a complete pipeline: face extraction, alignment, model training, and frame-by-frame compositing. Unlike consumer apps, these tools expose every hyperparameter, letting creators trade training time for output fidelity.

Pretrained weights for face synthesis, voice cloning, and lip-sync generation circulate freely across developer communities, stripping away the need for enormous compute clusters or curated datasets. These models rarely ship with built-in safeguards.

Watermarking and provenance tracking are absent unless deliberately added, and most repositories include no identity verification mechanisms. The default posture is unrestricted creation, with any guardrails added downstream by platforms or end users.

Training From Scratch vs. Using Pre Built Applications

The third dimension is a decision: build a model from scratch or use something pretrained. Training a custom deepfake model demands a high-end GPU, hundreds of thousands of aligned face images, and days to weeks of compute time. The output can be indistinguishable from genuine footage, tailored to a specific individual with precise control over expression and movement.

This is the approach behind the Arup wire fraud, in which cyberattackers constructed a synthetic executive convincing enough to pass on a live meeting, as documented by the World Economic Forum.

Pre-built applications and cloud services sit at the opposite end of the spectrum, trading fidelity for speed and delivering usable output in minutes while constraining the creator to built-in templates. Deepfake-as-a-service platforms bridge this gap aggressively, offering on-demand voice cloning and face synthesis through per-minute pricing that eliminates the GPU requirement and compresses a specialist operation into a transaction anyone can complete.

The same accessibility that lets defenders build realistic scenarios lets cyberattackers execute targeted impersonation campaigns. Adaptive Security's phishing simulations close that gap by putting employees inside the same tooling threat actors use.

Explore the platform

How Deepfakes Are Weaponized in Cybercrime and Fraud

When the generative adversarial networks and autoencoders that power deepfake creation fall into the hands of organized criminals, the result is immediately financial rather than experimental. Forbes' analysis of federal fraud data found U.S. deepfake-related fraud losses reached $1.1 billion in 2025, tripling from $360 million the year prior. The same face-swapping and voice-cloning techniques that produce convincing synthetic media in under an hour now back a multi-billion-dollar criminal enterprise operating across email, phone, and live video channels.

Executive Impersonation and Business Email Compromise

The most financially devastating deepfake cyberattacks combine synthetic audio and video with the psychological architecture of business email compromise. Cyberattackers no longer rely on a spoofed email alone; they reinforce the fraudulent request with a cloned voice call and, in sophisticated operations, a live deepfake video conference.

In January 2024, a finance employee at UK engineering firm Arup joined what appeared to be a routine video call with the company's chief financial officer and several colleagues. Every participant on that call was a deepfake.

The cyberattackers built the fake participants using publicly available video and audio harvested from earnings calls and conference presentations, the same OSINT-sourced training data that generative models require to produce convincing face swaps and voice clones. The employee authorized the transfers before discovering the deception.

According to the FBI Internet Crime Complaint Center's 2025 Internet Crime Report, released April 2026, business email compromise accounted for $3.046 billion in losses across 24,768 incidents, averaging roughly $123,000 per case, virtually all routed through manager-level approvers. What makes these cyberattacks so dangerous is how they exploit the same identity-verification instincts organizations have spent decades building. An email from the CFO confirmed by a voice call and a video call does not look like fraud; it looks like diligence.

Not every attempt succeeds, and the difference often comes down to a single verification habit. In July 2024, a Ferrari executive received WhatsApp messages that appeared to come from CEO Benedetto Vigna, describing a confidential acquisition and requesting urgent financial cooperation. A follow-up phone call used a cloned voice that closely matched Vigna's accent and speech pattern.

The executive grew suspicious and asked a personal question only the real Vigna could answer, drawn from a recent conversation the two had actually had. The caller could not answer and ended the call immediately, and the attempted fraud failed without any loss. That single out-of-band verification step, a question no cloned voice or scripted response could anticipate, is the exact behavior organizations now need to train at scale rather than leave to individual instinct.

Financial Fraud, Wire Transfer Scams, and Synthetic Identity

Beyond executive impersonation, deepfake technology enables categories of financial crime that do not require a live interaction to succeed. Synthetic identity fraud, where cyberattackers combine real and fabricated personal data to create fictitious identities, has been supercharged by generative AI's ability to produce realistic headshots and forged documents at scale.

Criminals use GAN-generated faces to open bank accounts and file fraudulent insurance claims. Because the identity is partially real and partially synthetic, it often passes automated verification checks that would flag a fully stolen identity.

Synthetic document fraud surged over 300% in the United States during the first quarter of 2025 alone, according to Sumsub's newsroom research on identity fraud trends, with deepfake-specific fraud attempts in that same U.S. window jumping 1,100%. The same autoencoder architecture that enables a face swap in a video call also powers the document forgery pipeline behind synthetic identity operations.

That crossover between video fraud and document fraud is precisely why financial institutions can no longer treat the two as separate risk categories.

Deloitte's Center for Financial Services projects that generative AI could drive total U.S. fraud losses to $40 billion by 2027, up from $12.3 billion in 2023. Wire transfer scams have evolved in parallel, moving from the single-vector CEO fraud email of a decade ago to coordinated assaults that deploy an AI-generated phishing email first, followed by a cloned voice call and, in high-value cases, a real-time deepfake video interaction.

The Psychological and Societal Impact of Deepfake Proliferation

The damage deepfakes cause extends beyond direct financial loss. As synthetic media saturates the information environment, it triggers what researchers call the liar's dividend: the ability for anyone caught in genuine audio or video evidence to dismiss it as fake. When a real recording and a deepfake become equally deniable, the shared evidentiary foundation institutions rely on begins to erode.

This erosion produces a creeping public exhaustion with determining what is authentic. When employees and consumers can no longer trust their own senses, the default response shifts toward skepticism of everything, including legitimate communications from employers and banks.

"Deepfakes don't just introduce falsehoods into an information ecosystem, they erode the very mechanisms by which societies construct shared understanding," said Dr. Nadia Naffi, Associate Professor of Educational Technology at Université Laval, in a UNESCO analysis of synthetic media's impact on public trust. Her research frames deepfake proliferation as a crisis of knowing itself, one that demands organizations rebuild verification into workflows rather than treating it as an individual skill.

For security teams, the practical implication is direct. The neural network architectures that produce deepfakes will continue improving in fidelity while decreasing in cost, and the cyber threat chain that begins with OSINT data collection and ends with a fraudulent wire transfer is now fully industrialized.

Multi-channel confirmation, email, call, video, used to signal legitimacy; now it is the cyberattack's biggest asset. Adaptive Security's Cloud Email Security closes the gap between what a channel confirms and what happened.

Book a demo

Are Deepfakes Illegal? Laws, Regulations, and Ethical Frameworks

Creating a deepfake is not inherently illegal in most jurisdictions; the act of generating synthetic media sits in similar legal territory to using photo editing software. What is criminalized, and increasingly so, is how that synthetic media is used: fraud, non-consensual intimate imagery, election interference, and impersonation. In 2026, lawmakers globally are closing the gap between what technology enables and what the law prohibits, though the regulatory landscape remains a fast-evolving patchwork.

U.S. Federal and State Deepfake Legislation

The most significant federal action arrived when President Trump signed the TAKE IT DOWN Act into law on May 19, 2025, criminalizing the knowing publication of non-consensual intimate imagery regardless of whether the content is authentic or AI-generated. The Act requires covered platforms to remove flagged content within 48 hours of a valid request, with the FTC empowered to levy civil penalties of up to $53,088 per violation. The first conviction under the law came in April 2026, when a 37-year-old Ohio man pleaded guilty to using AI to create and distribute synthetic intimate imagery of victims.

At the state level, 47 states have enacted deepfake-specific legislation as of mid-2026, according to a legislative database tracking AI-generated media laws. These laws generally fall into two categories: prohibitions on non-consensual intimate imagery, and disclosure requirements for political communications. As of May 2026, 30 states require AI-generated political advertisements to carry clear disclaimers, a patchwork that will face its first major test during the 2026 midterm elections.

International Regulatory Frameworks and Their Approaches

The European Union's AI Act includes Article 50 transparency obligations, taking full effect on August 2, 2026. The provision requires that AI-generated synthetic content be marked in a machine-detectable manner and that users be informed when interacting with an AI system. Rather than banning deepfakes outright, the EU approach mandates disclosure, placing the burden on platforms and deployers to label content rather than remove it.

India took a different path in early 2026 when its Ministry of Electronics and Information Technology amended its IT Rules to introduce the concept of synthetically generated information, imposing aggressive takedown timelines for government-flagged and high-risk content. The amendments also require labeling and metadata embedding for AI-generated content, and organizations operating in India should verify the specific takedown windows against the ministry's published amendment text.

"There's no putting this genie back in the bottle," said Dr. Hany Farid, professor at UC Berkeley's School of Information and a longtime authority on digital media forensics. "This is our new reality, and organizations are going to have to start thinking about how to put some guardrails on this technology before it ends up taking us somewhere nobody wants to go."

Content Provenance, Watermarking, and Ethical AI Development

The Coalition for Content Provenance and Authenticity, founded in 2021 by Adobe, Arm, BBC, Intel, and Microsoft, has emerged as the dominant technical standard for embedding cryptographically signed provenance metadata into digital media. Samsung's Galaxy S25 became the first consumer smartphone with native support for the standard, though its implementation applies content credentials only to AI-edited images rather than every camera capture.

But the coalition's standard has structural limitations. RAND Corporation, in its 2025 commentary Overpromising on Digital Provenance and Security, noted that the success of the ecosystem relies on end-to-end compliance across every image creation element, and in an open ecosystem the expectation that every participant will uphold designated standards is questionable.

Metadata is stripped when images are screenshotted, uploaded to social platforms, or format-converted, which are precisely the distribution chains where provenance matters most. The standard also certifies content history rather than truth; a deepfake produced by a tool that implements the standard will carry a technically valid manifest stating it was AI-generated, while a staged photograph taken with a compliant camera will carry an equally valid manifest saying it was captured in-camera.

Ethical guidelines for generative AI developers are coalescing around several norms: mandatory content credentialing at generation time, refusal to build tools designed specifically for non-consensual impersonation, and adversarial red-teaming against known deepfake misuse patterns. Legal clarity alone does not translate into employee preparedness, which is why the gap between what the law forbids and what employees can recognize in real time remains the operational frontier for most organizations.

Legal clarity around deepfakes does not stop a fraudulent wire request from reaching a finance team. Adaptive Security's Compliance Training keeps policy current while awareness training builds recognition skills law alone cannot provide.

Take a self-guided tour

How Deepfake Detection Works, and the Arms Race Behind It

Detecting deepfakes requires analyzing visual artifacts, audio inconsistencies, and behavioral anomalies across multiple signal categories, then feeding those findings into AI models trained on massive datasets of real and synthetic media. No single detection method is sufficient on its own; effective identification demands layered analysis across biological signals, pixel-level forensics, and contextual behavior.

Visual, Audio, and Behavioral Artifacts That Reveal Deepfakes

The most reliable deepfake detection starts with signals the human eye can be trained to spot, but at a granularity only machine analysis can consistently apply. Visual artifacts cluster around the face: inconsistent eye blinking, irregular pupil reflections, boundary artifacts along the jawline and hairline, and skin texture mismatches all betray synthetic media. Unnatural lighting and shadow directions that violate the scene's physics create additional forensic markers.

Audio-visual synchronization errors represent a particularly strong signal that single-modality analysis misses. When lip movements do not align precisely with phonemes in the audio track, the mismatch points to manipulation. The UK Department for Science, Innovation and Technology's 2026 deepfake detection market report identified multimodal detection, analyzing video and audio together rather than in isolation, as one of the most promising technical approaches for improving real-world accuracy.

Behavioral and contextual signals complement technical detection. A video call from a chief executive at 3 a.m. requesting an urgent wire transfer violates behavioral norms regardless of visual quality. Multi-channel attack patterns, where an email, voice call, and video meeting all apply simultaneous pressure, should trigger verification protocols even when each individual piece passes visual inspection.

AI Detection Tools, Training Datasets, and Techniques

Automated detection tools fall into two broad architectural categories. Passive detection analyzes media after creation using forensic artifact analysis. Active liveness detection verifies that a live subject is physically present during a real-time interaction, challenging the subject to turn their head or respond to randomized prompts rather than relying on artifact analysis alone.

These tools are trained on large-scale datasets containing over 100,000 real and synthetic videos across multiple generation techniques. The quality and diversity of training data directly determines detection performance. When a model is trained primarily on GAN-generated faces and then encounters diffusion-model outputs, accuracy collapses, a phenomenon researchers call cross-domain generalization failure and identify as the single largest technical barrier to operational deployment.

The Detection-Generation Arms Race: Who Is Winning and Why

Generators currently hold the advantage, and the gap is widening. Detection capabilities consistently lag behind generation advances; laboratory accuracy claims of 94% to 96% drop to 50% to 65% in real-world operational environments, where compression artifacts, varied lighting, and novel generation architectures all degrade detector performance simultaneously.

The structural asymmetry favors cyberattackers for several reasons. Generation models improve through normal market dynamics, since every GPU advance and open-source release immediately benefits deepfake creators. Detection models, by contrast, must be trained on examples of the latest generation technique after it appears, a reactive posture that guarantees lag.

Cyberattackers can also test their outputs against known detection tools before deploying them, iterating until the deepfake passes, while defenders cannot pre-test against generation methods that do not yet exist.

This brittleness means detection tools are permanently playing catch-up against an adversary that innovates on its own timeline. Closing the gap requires a shift from pure forensic detection toward layered verification, combining automated detection tools with behavioral verification protocols and out-of-band confirmation for high-risk requests.

Detection accuracy that looks strong in a lab drops sharply once a deepfake reaches an unpredictable channel. Adaptive Security pairs behavioral signals with realistic phishing simulations so verification never depends on one layer.

Explore the platform

Building Organizational Resilience Against Synthetic Media Threats

According to Verizon's 2026 Data Breach Investigations Report, the human element was present in 62% of confirmed breaches. Knowing how synthetic media is created is not an academic exercise against that backdrop; it is the operational foundation of every effective defense.

When organizations understand that synthetic media is sourced from publicly available data, requires near-zero technical skill to produce, and can be deployed within minutes, the defensive posture shifts from passive compliance to active behavioral readiness. Most employees have never been told how AI deepfakes are created, which means they have no framework for recognizing one in progress.

Why Understanding Deepfake Creation Directly Informs Better Defense

The mechanics of deepfake creation reveal exactly where organizational vulnerabilities live. Generative AI models train on publicly accessible material: conference recordings, podcast interviews, and earnings call transcripts all feed the pipeline that produces convincing synthetic replicas.

When employees understand this pipeline, open-source intelligence hygiene transforms from an abstract IT policy into a personal security behavior. An executive who knows a keynote recording can be harvested and fed into a voice cloning tool within minutes is far more likely to limit what gets posted publicly.

According to the National Cybersecurity Alliance's 2025-2026 Oh Behave! The Annual Cybersecurity Attitudes and Behaviors Report, 52% of employed participants reported receiving no training on the security or privacy risks of AI tools, despite 65% now using AI and 43% admitting to sharing sensitive work information with AI tools. This gap concentrates risk precisely where visibility is lowest, and it applies as much to how employees respond to AI-generated threats as to how they use AI tools themselves.

Training Employees to Recognize and Respond to Synthetic Media Threats

Effective training against synthetic media does not rely on teaching employees to spot visual artifacts or audio glitches, since those cues disappear with every model iteration. Instead, it builds behavioral reflexes that operate regardless of how convincing the fake becomes.

Voice verification protocols form the first line of defense. Employees who handle financial transactions must be trained to confirm any verbal instruction through a second, out-of-band channel, such as a Slack message, a text to a known number, or a pre-established code word. This habit is not intuitive and must be rehearsed repeatedly through structured vishing simulation programs before it becomes reflexive.

Multi-channel authentication for financial requests adds a procedural backstop. A wire transfer request that arrives via email and is confirmed by a voice call should still trigger a third verification step, never two channels alone, since deepfake cyberattacks increasingly coordinate across email, voice, and video simultaneously. Real-time recognition training shifts the employee's mental model from asking whether something looks fake to asking whether the request itself makes sense, teaching employees to flag urgency manipulation and deviations from standard payment workflows.

Continuous Security Awareness as the Foundation of AI-Era Defense

Annual compliance training was built for a world where threats evolved slowly, but deepfake tools improve monthly. Continuous awareness means phishing simulation cadences measured in weeks rather than years, training modules that update as cyberattackers adopt new tools, and risk scoring that tracks which departments face the highest synthetic media exposure.

As NIST computer scientist Julie Haney and University of Maryland Associate Professor Wayne Lutters concluded in their peer-reviewed analysis published in Computer in October 2020, compliance metrics do not tell the whole story and fail to measure whether a program produces sustained change in employee attitudes and behaviors. That finding, though written before the deepfake era, applies directly to how organizations should evaluate readiness against synthetic media: completion rates say nothing about whether an employee would actually pause before wiring funds to a convincingly cloned voice.

The underlying principle is straightforward. Deepfake creation has become fast, cheap, and accessible, and the only defense that scales at the same speed is a workforce trained to recognize the attack pattern rather than the artifact. That requires moving from awareness as a compliance checkbox to awareness as an operational capability, one that is practiced, measured, and continuously refined.

See How Adaptive Security Trains Employees to Recognize Deepfake Threats

Organizations that understand how are AI deepfakes created still need a way to translate that knowledge into employee behavior that holds up under pressure. Adaptive Security's cybersecurity awareness training platform closes that gap by training employees to recognize the specific visual, audio, and behavioral cues that distinguish real communications from AI-generated deepfakes, turning a workforce into an active detection layer rather than a passive audience for annual compliance modules.

The cybersecurity awareness training platform's phishing simulation engine runs deepfake-specific video and voice scenarios alongside traditional email cyberattacks, so the multi-channel pressure that made the Arup fraud successful becomes a rehearsed pattern rather than a first-time shock. Cloud Email Security adds a detection layer purpose-built for AI-generated business email compromise, catching the spoofed executive emails that typically open a deepfake cyberattack chain before an employee ever sees them, while Compliance Training keeps jurisdiction-specific regulatory obligations current as deepfake legislation continues to expand.

Results compound because every detected threat and every phishing simulation outcome feeds back into individual risk scoring, letting security teams direct reinforcement training toward the departments facing the highest synthetic media exposure. Continuous, threat-informed cybersecurity awareness training replaces the static, once-a-year model that cannot keep pace with monthly improvements in deepfake tooling.

Most organizations are still training employees to spot email typos while cyberattackers rehearse full video impersonations. Adaptive Security builds the multi-channel readiness that closes that gap.

Take a self-guided tour

Frequently Asked Questions About How Are AI Deepfakes Created

How Long Does It Take to Create an AI Deepfake?

Creation time ranges from under 90 seconds on consumer-grade applications to roughly 25 minutes for a convincing 60-second video using freely available tools, according to the World Economic Forum. A brief voice sample can achieve 85% accuracy in seconds, while real-time face-swapping tools add well under 100 milliseconds of latency, making live deepfakes effectively instantaneous during an active call rather than something a cyberattacker prepares hours in advance. The wide range matters operationally: a five-minute vishing call and a 45-minute video conference impersonation both fall within a single work break, which is why verification protocols cannot assume a cyberattacker needs meaningful lead time to strike.

Can Deepfakes Be Generated in Real Time During a Live Video Call or Phone Conversation?

Yes. Open-source tools intercept a webcam feed, perform face-swapping locally on a GPU, and route the altered video into Zoom, Google Meet, or Microsoft Teams with minimal added latency. On the audio side, real-time voice conversion systems can alter a speaker's voice during a live phone call with latency low enough to sustain a natural conversation.

Some video conferencing platforms have responded by building deepfake detection directly into their software, though coverage remains inconsistent across providers and does not extend to standard phone lines.

What Is the Difference Between a Deepfake and a Shallowfake?

A deepfake is media generated or substantially manipulated by an AI model trained on a target's likeness or voice, while a shallowfake, also called a cheapfake, uses conventional non-AI editing such as slowed footage or clips spliced out of context. The practical distinction for a security team is response strategy: shallowfakes typically yield to source verification and reverse image search, while deepfakes require behavioral verification protocols because their artifacts exist at a pixel and waveform level most employees cannot consciously detect, even with training.

How Much Source Material Is Needed to Create a Convincing Deepfake?

Requirements vary sharply by quality tier. A single photo and a few seconds of audio can produce a recognizable but flawed deepfake using consumer apps, while photorealistic, high-fidelity deepfakes demand thousands of multi-angle images plus hours of video.

The trend across 2025 and 2026 has been toward requiring less, not more. Newer one-shot voice generation tools now claim usable output from a handful of seconds of sample audio, which means the floor for a usable cyberattack keeps dropping even as the ceiling for photorealism keeps rising, a combination that narrows the gap between casual social media misuse and targeted executive impersonation.

Are AI Deepfakes Illegal to Create or Distribute?

Deepfake creation is not inherently illegal, but specific uses are criminalized under a rapidly expanding patchwork of laws. What determines legality is use rather than creation: the same technical pipeline that produces a harmless video filter can produce a federal crime under the TAKE IT DOWN Act if the output depicts nonconsensual intimate imagery, or a wire fraud charge if used to impersonate an executive. For organizations, that distinction matters less than preparedness, since the difference between an averted incident and a costly one usually comes down to whether an employee was trained to verify before acting, not whether the underlying deepfake technology was itself illegal to create.

Key Takeaways

• How AI deepfakes are created follows a consistent pipeline: data collection, preprocessing, model training, generation, and post-processing, and each stage represents a point where organizational awareness can disrupt the cyberattack.
• Understanding how are AI deepfakes created is what allows security teams to recognize the specific artifacts and behavioral patterns that distinguish synthetic media from genuine communication.
• Voice cloning and face-swap technology, core to how AI deepfakes are created, now require minimal source material, meaning nearly every executive with a public presence has already published enough data for a usable clone.
• Detection tools remain locked in a permanent arms race with generation technology, which means organizational defense cannot rely on detection alone and must build behavioral verification into standard workflows.
• Multi-channel verification is the single most reliable defense against the fraud that stems from how are AI deepfakes created, regardless of how convincing the synthetic media becomes.
• Continuous, realistic cybersecurity awareness training that includes deepfake-specific scenarios builds the recognition instincts that annual compliance training cannot.

No employee can be expected to catch a deepfake under real pressure without ever having rehearsed one. Adaptive Security builds that readiness through realistic, multi-channel phishing simulation programs run on a regular cadence.

Book a demo