Deepfake Video Detection Explained: How to Spot Them and How to Stay Safe

Following the February 2024 Arup deepfake attack, awareness of the repercussions of a face-swap attack has pushed deepfake video detection towards becoming a forensic discipline. A competent attacker is now able to generate a broadcast-quality video face-swap in under fifteen minutes using widely available consumer software. This puts a lot of pressure on detectors that need constant updating to stay apace with attackers.

This blog walks you through the deepfake video detection methods that a defensible stack must have in 2026, how each layer works behind the scenes, where they break, and how to stack them so no single failure leaves your organization exposed.

What is a Deepfake Video?

A deepfake video is a manipulated or synthetic representation in which a person’s face, expressions, body, or complete identity has been generated or altered using AI.

There are four attack classes that most security teams encounter:

• Face-swaps: Where a real person’s identity is replaced with another across frames
• Re-enactment or expression-swaps: Where the target’s identity is real but their lip movement and expressions are controlled using a driving video
• Full-body puppeteering: Where the entire body of a target is animated by a source performer
• Neural talking heads: Where a single-referenced image is animated using a target audio track.

Each of these classes has distinct architectures that are generally one of three families: a generative adversarial network (GAN), a neural talking-head system, or a diffusion model. Each class also has its unique signature, therefore requiring a combination of targeted detection techniques to unravel.

What is Deepfake Video Detection?

The technical discipline of identifying whether a video is wholly synthetic, partially manipulated, or genuine is called deepfake video detection.

The output of a modern detector is not a binary label but a probability score. The artifact map tells analysts which frames, which regions of a face, and which frequency bands triggered the verdict.

Detectors produce probabilistic outputs because the right confidence threshold depends on how they are used. An identity verification workflow requires a different threshold than a newsroom publication workflow. Any vendor selling a verdict without a confidence score along with an artifact map and an evaluation benchmark is most likely selling you a demo.

Why Deepfake Video Detection is Harder in 2026 Than It Was Before

There are four reasons that explain why it has become harder for defenders to use video deepfake detectors.

• Generator Proliferation: In 2026, face-swap attack ecosystems include dozens of commercially available neural talking-head services, open diffusion models that were fine-tuned for face manipulation, and video generation platforms that ship a new model every quarter.
• Diffusion-model Dominance: Diffusion architectures produce frame-level output that lacks many of the GAN-era giveaways detectors were built to read. Artifacts such as frequency-domain fingerprinting and up-sampling patterns that were exploited earlier are either completely absent or significantly weaker in diffusion output.
• Compression and Channel Loss: Most deepfake videos reach the victim over a compressed channel, such as WhatsApp forwards or a Zoom call. Codec compression is known to strip fine spatial and frequency details that most new generation classifiers depend on.
• Cross-modal Attacks: Incidents like the Arup case, the wave of executive-impersonation attacks, and the 2022 fabricated Zelenskyy capitulation video, all involved a combination of synthetic videos, cloned voices, spoofed calendar invites, or a plausible pretext. Detectors able to read videos in isolation were able to detect video-only attacks, while multimodal detectors like ours were able to detect the audio as well. We cover this in greater detail in our guide to deepfake audio detection methods.

How Deepfake Video Detection Works

Every deepfake video detection method answers one of three underlying questions:

1. Does the content itself look fake?
2. Where did this file actually come from?
3. Is there a real, living human physically present in this stream at this moment?

A strong detection system requires following the methods listed below to answer all three questions.

Spatial Artifact and Blending Seam Analysis

This is one of the most useful and oldest ways to detect a deepfake. It tries to detect the area where a fake face has been placed over the original head. Face-swapping tools use a blending mask to combine generated faces with real video feeds, leaving room for imperfections.

How it works: Vision transformers or convolutional neural networks are trained to compare real and synthetic faces. They pay close attention to the edges of the face where a blended mask is most likely to show. Some specialized detectors can also examine the image in the frequency domain instead of looking at individual pixels. Such detectors come in handy when the video has been compressed.

What it catches:

• Color and brightness differences around the hairline, jaw, and ears.
• Level of detail between the synthetic face and the surrounding area where it has been placed.
• Alignment errors when the head moves.
• Minute skin texture differences on the synthetic face and the neck.

Where it breaks: If a video has undergone heavy compression or resizing, the tell-tale artifacts that these detectors look for may pass scrutiny. Diffusion-based face manipulation systems can circumvent this problem by generating the entire frame instead of using a blending mask to place a fake face over a real one.

Temporal Consistency and Inter-Frame Coherence

In a real-life video, a person’s face moves continuously from one frame to the next. Deepfake videos often produce inconsistent movements that can give them away.

How it works: AI tools analyze multiple frames at once rather than looking at a single image to learn what natural movement looks like, and be able to distinguish between synthetic and real frames. More modern deepfake generators are now able to avoid obvious artifacts such as motion blur moving against the background.

What it catches:

• Small changes in eye position, mouth movement, and nose twitches between frames.
• Inconsistent blinking
• Jerks in head movements, especially when viewed from frame to frame
• Facial expressions that lack small movements that are normally seen in a real face.

Where it breaks: Videos under two seconds throw off the detectors since they do not contain enough information. Small movement patterns that are necessary for optimum detection are almost completely removed when videos are heavily compressed.

Physiological Signal Detection

Remote Photoplethysmography (rPPG) is used to detect the slight variations in skin tone that a real person’s heartbeat causes as blood flows under the skin.

How it works: The detector examines color changes across the face, such as the forehead, cheeks, and jaw, as natural blood flow leaves characteristic variations in skin tone. It filters out irrelevant changes and looks for a pulse signal within a normal heart-rate range. It then matches the signal strength with the expected blood flow to determine if the face is real or synthetic.

What it catches:

• Detects a human heartbeat through the camera
• Detects anomalies in the pulse of the person on camera. It should ideally vary naturally across different parts of their face.

Where it breaks: rPPG works best in clear, well-lit, and relatively stable video feeds. Videos taken on patchy mobile networks or poor-quality webcams, in low-light conditions, involving moving subjects, or even with heavy compression reduce the reliability of the detectors significantly.

Biometric and Behavioral Analysis

People express themselves in their own subtle ways. When they talk and express themselves, they move in varying ways. During a video session, these behaviors remain fairly consistent for a real person, but AI-generated faces find it difficult to reproduce the movements accurately.

How it works: AI models, when trained to observe and learn the behavior, movements, and hand gestures of a specific target, such as a CEO, can identify unusual patterns. When a synthetic video surfaces, an executive protection system can flag it if the behavior does not match.

What it catches:

• Minute facial changes that happen before or after a specific expression.
• How a person moves their eyes and changes their focus.
• How long and how often they blink.
• Small individual habits, such as how they move their lips before speaking.

Where it breaks: The process of creating a reliable behavioral profile for a new executive requires enough data that needs to be updated over time as their behavior patterns change naturally.

Frequency-Domain and Spectral Analysis

Deepfake video generators leave behind subtle signals in images that are easier to see when they are analyzed by their frequencies instead of the visible pixels. Frequency analysis remains a useful method to detect deepfakes, though the patterns keep changing as generation technology improves.

How it works: A detector runs an FFT or DCT classifier that runs across each frame after a quick denoising pass. This allows the detector to read the spectral and frequency distribution against a baseline of what a natural camera output might produce.

What it catches:

• Energy patterns at specific frequencies that are introduced by GAN generators
• Over-smoothed low-frequency content
• Photo Response Non-Uniformity (PRNU) that points to camera sensor noise that may be missing or broken in manipulated frames of a synthetic video.
• Electric Network Frequency (ENF) that detects electrical grid flickers that match the local power grid at the time and location of capture.

Where it breaks: When videos are compressed aggressively, the process flattens out its signature frequency. A skilled attacker can use PRNU forgery to bypass a detector by adding a sensor noise to the deepfake output post-processing.

AI Fingerprinting and Model Attribution

All generative models leave characteristic fingerprints in their output.

How it works: AI fingerprinting cuts through the noise distribution, sampling residues, and frequency spectrum, to read these statistical traces and attributes the generated video back to the model that produced it.

What it catches:

• Artifacts that can be attributed to specific generative models.
• Watermarks embedded in the video.

Where it breaks: A December 2025 University of Edinburgh study evaluating 14 fingerprinting methods against 12 image generators found that fingerprint removal succeeded over 80% of the time when an attacker had full access to the fingerprinting model, and over 50% with only black-box access. Forgery, where an attacker stamps content with another model’s fingerprint, proved harder and varied by target model.

Media Authentication and Provenance

This method checks for a verifiable record of where the content came from, instead of inspecting the pixels themselves.

How it works: The Coalition for Content Provenance and Authenticity (C2PA) binds a cryptographically signed manifest to a media file at the point of capture or edit, recording the device, the software, and every subsequent modification.

What it catches:

• Mismatch against the provenance manifest.
• If the content has any signs of tampering after signing.

Where it breaks: Provenance only proves the origin of the content. Therefore, a bad actor can produce a technically valid manifest for synthetic content. Most genuine content also carries no credentials at all, so the absence of a manifest proves nothing on its own.

Real-Time Liveness and Injection Attack Detection

These are guided by two separate, and independently certified controls that achieve two different outcomes.

Liveness detection. It detects if a real human is physically present in front of the camera.

What it catches: It detects the presence of screen replays, silicone masks, and printed photographs.

Where it breaks: Liveness detectors are unable to see an attack that has been injected into a live video stream, because it does not pass through the lens.

Injection attack detection. This control is used to detect if an attacker has pushed any deepfake video into a live video feed.

What it catches: Detects if any virtual camera drivers, emulators, or manipulated SDK calls have been added directly into the media stream.

Where it breaks: It breaks at the boundary where the software trusts the data pipeline instead of verifying the physical devices or origin of the input.

Anatomy of a Deepfake Video Attack

It is always better to decide which detection method suits your enterprise when you are aware of the attack schematics that the detectors will need to detect. Here is a list of the typical enterprise-facing video attack techniques:

1. Reconnaissance: An attacker typically selects a target that has wire-transfer authority or system access. They then scour the Internet for reference footage to help them generate a clean frontal video that can be used in a face-swap attack. For some high-profile assets, attackers carefully maintain reference libraries that have footage exfiltrated from earlier breaches.
2. Training the Model: The next step is to generate an open-source face-swap or neural talking head model using the reference footage. The neural heads/face-swaps are trained using consumer GPU rentals that cost a few dollars per model.
3. Developing the Pretext: Deepfake attacks usually build up slowly through spoofed calendar invites, compromised email threads, or a malicious Slack message that sets the context for the video call. The Arup deepfake video call incident reportedly escalated following a targeted phishing email. The email set the narrative for the video call before the call was ever initiated.
4. The Delivery: An attacker either uses a live face-swap on their system to join a live video call, or injects a pre-recorded video directly into the meeting platform’s media stream. On the call, the synthetic executives request a wire transfer, an approval for a policy correction, or requests a credential reset.
5. Exfiltration: If the attack goes through and the target acts, the attackers drop the call and route the wire transfer through a multi-route chain so that by the time the target realizes they have been compromised, the money has been exfiltrated along with any other sensitive data.

Warning Signs of a Deepfake Video: What to Look For

While modern generation models have been able to close most of the obvious giveaways, the signals listed below are still useful supporting evidence for a trained pair of eyes and should never be treated as an authentication protocol or the sole basis for a verdict.

• Face and Hairline: A mismatch in skin texture around the face, the neck, or close to the hairline; blurred or shifting hairline or facial edges as the head moves; asymmetrical ears; and skin that doesn’t seem to match the age of the person are all telltale visual cues of a deepfake video.
• Eyes and Gaze: Sometimes, the reflections in the eyes do not match the lighting in the room, pupils don’t dilate or contract as the scene brightens or dims. Generated videos also have inconsistent blinking and drifting gaze.
• Mouth and Lip-Sync: Teeth and lip movements that lead or lag the audio consistently, or a tongue that disappears at the back of the mouth, or the absence of breath sounds at the start of a sentence.
• Lighting and Shadow: Shadows that do not match the direction of the light on the neck or upper body, highlights on the skin or glare on the glasses that stay static while the head moves.
• Motion and background: A head that moves independently of the shoulders, background objects that stretch or warp near the boundary of the face region, and motion blur in the frame that doesn’t add up.
• Color and Compression: When color bands appear only on the region of the swapped face, or a subtle rectangular frame seems to demarcate video clarity and quality around the face across multiple frames, it may be a sign of a deepfake video.

Where Deepfake Video Detection Sits in Your Enterprise Stack

Since deepfake video detection is a set of controls that are placed at a surface where a video-based decision is made, it is essential for enterprises to know where the detectors sit in your stack:

• KYC and Identity Verification: Workflows that require selfie-based onboarding are the most exposed surfaces since liveness detectors are unable to catch a possible injection attack. Such workflows require injection-attack detection certified against CEN/TS 18099 and conformant to NIST SP 800-63-4.
• Executive Video Approvals: Enterprises where a video call can be used to authorize financial transactions, or a credential reset, or policy alteration, require a mandatory out-of-band confirmation process in place.
• Brand Protection: A video showing an impersonated executive or vendor can crash company stock in minutes. A controlled monitoring feed that runs artifact-based forensic screening on video content that mentions or features your executives, can be paired with C2PA provenance so recipients can verify authenticity.
• Information Integrity: Posts on social media that experience a sudden uptick in reach should be paired with provenance verification, physical signal analysis, and automated screening.
• Law Enforcement and News: Editing workflows, forensic analysis that include rPPG and PRNU suites, and audit trails of the original media should be backed by provenance-aware detection verdicts and confidence scores.

How to Stay Safe from Deepfake Videos

Enterprises looking to stay safe from deepfake videos must deploy detection at every surface where action is authorized on a video call. Each surface requires its own class of controls tuned to catch the discrepancies that the layer specializes in.

We recommend organizations to layer artifact analysis, temporal analysis, physiological detection, behavioral profiling, provenance verification, and injection-attack detection since each has its own blindspots.

It is also advised that detection accuracy be treated as something that decays with time as newer generation models are introduced and requires regular retraining and benchmark testing against in-the-wild datasets like Deepfake-Eval-2024 and specific industry-recognized evaluation metrics, such as Attack Presentation Classification Error Rate (APCER), Bona Fide Presentation Classification Error Rate (BPCER), and Equal Error Rate (EER) for biometric systems and Area Under the ROC Curve (AUC) for artifact detection.

Raising awareness amongst enterprise employees and running tabletop exercises against deepfake schematics helps strengthen incident response capabilities. Thus, it is crucial for companies to teach their executives the pretext patterns and out-of-band confirmation protocols that attackers use.

Where Diopter Fits

For any enterprise, the ability to distinguish a real video from a deepfake consolidates trust, and therefore should be a mandatory inclusion in their security budget.

Our multi-layered video deepfake detection system reports independently, retrains against updated generators every quarter, and reflects the current threat model in play.

Review our solutions to ensure which of the method families your current stack actually covers, where the gaps lie, and how quickly your detectors would be able to detect an attack.

FAQs