Free Deepfake Detector

How deepfake detection works

A deepfake is a real face replaced by a generated one, and every method of doing it leaves something behind.

The face-swap architectures leave the clearest trace. DeepFaceLab, FaceSwap and the tools built on them all run generated pixels through a decoder that has its own frequency signature, and that signature persists through the blend into the surrounding frame. GAN fingerprinting looks for it directly. It survives resizing and a round of re-compression, which is why it still works on an image someone screenshotted off a phone.

Facial geometry catches the blend itself. A swapped face has to be warped onto a head that was never shaped for it, and the seam shows up as inconsistency between the geometry of the face and the geometry of the skull, ears and jaw around it. Ears are the classic failure: they belong to the original person, and they rarely match.

PRNU is the one that is hardest to defeat. Every camera sensor has a unique noise fingerprint that appears across the entire frame of a genuine photograph. A swapped region carries a different fingerprint, or no sensor fingerprint at all, because no sensor made it. The face and the background disagree about what camera took the picture.

In video, temporal analysis adds a dimension that still images do not have. Each frame is generated independently, so consistency between frames is something the model has to achieve rather than something it gets for free. Flicker at the blend boundary, jitter in facial landmarks, and lighting that shifts by a fraction between frames are all artifacts of that.

Every signal returns its own status and its own explanation. When they disagree, you see the disagreement rather than an averaged-away verdict, which matters because the disagreement is usually the interesting part.

Why this got urgent

The cost of making a convincing fake collapsed, and the cost of checking one did not.

Voice cloning now needs about three seconds of reference audio. Deepfake video can be produced for less than a dollar. Meanwhile verification still requires either expertise or a tool, and most people encountering a suspect image have neither to hand.

The documented damage follows the money. A finance employee at Arup authorised a 25 million dollar transfer after a video call in which every other participant, including the company's CFO, was synthetic. Romance scams built on AI profile photos ran to 1.3 billion dollars in reported US losses in 2025. Deepfake-enabled fraud overall reached 12.5 billion.

What has changed most recently is latency. Real-time face swap runs on a consumer GPU inside the frame budget of a video call, which removes the last piece of advice everyone was giving: that you could trust a live conversation. You cannot, and the section below covers what to do about that instead.

None of which makes detection a solved problem. It makes it a necessary one.

Detecting deepfakes in video

Video is a harder target than a still, and a more informative one.

Upload an MP4 or MOV up to 100MB. The engine extracts key frames, runs the full six-signal analysis on each, and then does something a still image cannot support: it compares frames against each other.

That temporal pass is where video-specific generators give themselves away. Sora, Veo, Kling and Runway Gen-3 each produce frames that are individually plausible and collectively slightly inconsistent, because temporal coherence is something the model approximates rather than obeys. Watch a generated clip frame by frame and you find edges that breathe, texture that reorganises between frames, and shadows that drift independently of the object casting them.

Face-swapped video fails differently. The underlying footage is real, so the background is temporally perfect while the swapped region is not. The mismatch between a stable scene and an unstable face is itself the signal, and it is one of the more reliable ones we have.

Two practical notes. Heavily compressed video, anything that has been through a social platform more than once, degrades the frequency evidence and lowers confidence. And a short clip gives the temporal pass less to work with, so a three-second video is a weaker test than a thirty-second one.

Real-Time Deepfake Filters: The New Frontier of Live Call Fraud

Real-time deepfake filters, tools that replace a person's face during a live video call, represent a fundamentally different threat than pre-recorded deepfake video. Applications like DeepFaceLive, Avatarify, and commercial products built on them can run on consumer GPUs and apply a face swap in under 30 milliseconds, well within the latency budget of a Zoom or Google Meet call. The victim sees what appears to be a real person speaking naturally.

Real-time deepfakes are detectable but require different analytical techniques than static images or pre-recorded video. The most reliable signals are: edge artifacts around the face boundary where the synthetic overlay meets the real background, reduced facial texture resolution compared to the rest of the video frame, an absence of micro-expression variation that real faces exhibit subconsciously, and unnatural blink patterns - most real-time deepfake models struggle to replicate the precise timing of human blinking.

FauxLens analyzes recorded video from calls (a screen recording of a Zoom session, for example) and identifies these real-time filter artifacts. For live detection during an ongoing call, the most practical countermeasure remains asking the other person to perform spontaneous tasks: hold a specific object, look directly left or right quickly, or read a randomly generated word from a sheet of paper. Current real-time deepfake systems cannot process these rapid, unpredictable requests reliably.

Deepfake Detection Use Cases by Industry

Deepfake detection has become operationally necessary across multiple industries, each with different threat profiles and stakes.

HR and hiring: The FBI issued a public warning in 2022 and updated it in 2025 documenting North Korea-linked operatives using deepfakes to fraudulently obtain remote employment at US technology companies. The goal is insider access to proprietary systems and source code. HR teams without deepfake screening in their video interview process are exposed to this threat. FauxLens screens recorded interview clips and profile photos for face-swap and synthesis artefacts, and returns the specific findings behind each verdict so a hiring team can see what was flagged rather than acting on a bare score.

Journalism: Deepfake images of public figures in fabricated situations are designed to create publishable controversy. In 2025, synthetic images of political figures were shared during two separate national elections before forensic teams identified them as AI-generated. Newsrooms integrating FauxLens into their photo intake workflow catch these before publication.

Banking and KYC: Identity verification for financial account opening increasingly uses video selfies and liveness checks. Criminal organizations have developed AI tools specifically designed to defeat liveness detection with synthetic video. Banks adding a second-layer forensic check on submitted verification video reduce successful synthetic identity fraud attempts.

Dating platforms: Romance scammers generate consistent AI personas across multiple photos using Midjourney or Flux. These synthetic identities cannot be caught by reverse image search because they have never been photographed. FauxLens API integration at the photo upload stage gives dating platforms the ability to flag potential synthetic profiles before they reach other users.

Insurance: AI-generated damage documentation (fabricated photos of fire damage, vehicle accidents, and property destruction) is submitted to claims departments at scale. Insurers using forensic AI detection at the claim intake stage have reduced AI-fabricated claim payouts significantly.

Legal proceedings: Courts in multiple US jurisdictions have encountered AI-fabricated photographic evidence submitted in civil cases. Several states have enacted rules requiring disclosure when AI tools were used in evidence creation, but detection capability is necessary when disclosure is not forthcoming.

How to Protect Yourself From Deepfake Fraud

Protecting yourself from deepfake fraud requires both detection tools and behavioral practices, because no single countermeasure is sufficient on its own.

For video calls and online relationships: establish a video-based liveness challenge before trusting any new relationship formed online or via professional platforms. Ask the person to hold up a handwritten code you provide in real time, or to quickly turn their face in a specific direction. Real-time deepfake systems cannot respond to arbitrary spontaneous visual tasks reliably. Request multiple unscheduled video calls at different times - deepfake actors frequently avoid or delay these.

For profile photos on dating apps and social platforms: upload the photo to FauxLens before emotionally investing in the relationship. AI-generated profile photos are the most common entry point for romance scams. Visual inspection is insufficient, the latest Midjourney and Flux outputs fool the human eye more than 70% of the time.

For submitted-video verification: record the clip and submit it to FauxLens for forensic analysis. Cross-reference the candidate's stated identity against government-issued ID through a liveness-based identity verification service. Require a live unscheduled call in addition to any pre-recorded submission.

If you suspect you are the victim of deepfake fraud and money has been transferred: contact your bank immediately and request a transaction recall. File a report with the FBI Internet Crime Complaint Center at ic3.gov and with the FTC at reportfraud.ftc.gov. Save all evidence (screenshots, recordings, messages) before contacting the platform where you met the person, as account deletion can be rapid once a scammer knows they have been identified.