Deepfake Detector

Historically, a “deepfake” meant a video face swap — a real person’s face replaced with someone else’s in moving footage. Today the word is used loosely for any synthetic depiction of a person, including fully generated still portraits that depict someone who never existed and AI-altered photos of people who do.

That is why this page overlaps with face-detection language. Whether you searched for a deepfake detector or an AI face detector, the underlying need is the same: decide whether a face in a still image is likely synthetic before you extend trust, access, or money.

The boundaries stay honest here: this page is for still images. If your concern is video — a suspicious call, a circulating clip, a face-swapped recording — you need video-focused forensics, because moving footage carries temporal artifacts that a photo checker never sees.

Romance and pig-butchering scams: polished, model-like portraits on brand-new accounts that move conversations toward money quickly. The portrait is often the only asset the account owns.

Fake recruiters and HR contacts: professional headshots attached to offers that arrive unsolicited, followed by requests for personal data, “equipment deposits”, or off-platform onboarding. Impersonated executives: a synthesized portrait of a known leader attached to urgent payment or credential requests, aimed at employees who will not question authority.

Marketplace sellers and review sections: borrowed or generated faces lending credibility to listings and five-star profiles. The common thread is a face doing trust work for an account with no other history — which is exactly the pattern a portrait check is built to interrupt.

Step one: run the portrait through the checker above and note the score. Step two: demand a live signal that a static image cannot satisfy — a video call, or a selfie holding a specific number of fingers or a written word you chose at that moment. Fully generated faces cannot respond to unscripted requests.

Step three: reverse-search the portrait. Synthetic faces sometimes surface across multiple unrelated profiles, which is close to conclusive. Step four: weigh the account’s context — creation date, friend or follower patterns, posting history, and whether the story behind the contact makes sense.

Only after those steps should you act on identity: block, report, or escalate inside your platform or organization. When the stakes involve money or safety, preserve the evidence — screenshots with URLs and timestamps — before accounts are deleted, and report through the platform’s impersonation or fraud channels.

A photo checker and a video deepfake detector face different artifact families. Still portraits carry texture, boundary, and lighting cues inside one frame. Video adds temporal signals — flicker between frames, unstable facial landmarks, blending errors at the face boundary — that only appear over time.

This tool covers the still-image case, which is where most first contact happens: profile pictures, headshots, and shared photos. If a video is your concern, use it only on extracted frames as a weak signal, and rely on video-specific forensics for anything consequential.

Voice is the same story one channel over: cloned audio convincing enough to imitate a family member or executive needs procedural defenses — agreed code words, call-backs to known numbers — rather than content analysis. Deepfake defense is layered because the attack surface is layered.

Every static-image technique eventually hits the same wall: a still photograph cannot prove a person is present at this moment. Liveness challenges break that wall by requiring a response the file cannot contain. Ask the person to hold up a hand showing a specific number of fingers, say a phrase you choose at that moment, turn to a light source, or move their head through an angle while keeping their face visible. The response must be unscripted, and it must be live.

Why this beats the score: a generated face can satisfy a static photograph completely — it was generated to be photographically consistent. It cannot improvise. The gap between “looks like a plausible person in a still frame” and “responds correctly to an unscripted instruction you invented two minutes ago” is exactly the gap that separates synthetic identity from a real person, and only the second one is decisive.

Build the challenge into your process rather than treating it as an interrogation. Randomise the requested gesture. Ask for a specific phrase rather than a yes-or-no. Combine it with a document check, since a document and a live face appearing together is much harder to synthesise than either alone. Keep the friction low for genuine users and high for attackers — the goal is not to inconvenience honest people.

Then weigh the account context, which is free to check and often decisive: creation date, posting history, whether the account suddenly began contacting people in a different demographic, whether the story behind the contact makes chronological sense, and whether the same portrait appears on unrelated profiles. A face that has no history behind it is telling you something regardless of what the score says.

Pick thresholds that match your risk. A common shape: low scores and low-risk actions proceed automatically; mid-range scores queue for human review; high scores hold the action until identity is confirmed another way. Publish the policy internally so reviewers apply it the same way every time.

Never automate punishment. Scores drift with filters, lighting, camera quality, and skin tone, which is why an automatic ban on a threshold alone creates unfair outcomes. The score should route attention, not deliver verdicts — a human makes the call whenever consequences attach.

Keep a record of consequential decisions: the score, the evidence gathered, the reviewer, and the outcome. That record protects real users who are wrongly flagged, helps you tune thresholds over time, and gives you a defensible position if a decision is later challenged.