Ilia Zavialov deepfake fraud glossary: plain meanings for the words used around forged video
Once a disputed clip starts circulating, the vocabulary around it multiplies quickly. Platforms answer with policy language, detection vendors answer with percentages, and journalists reach for whichever word sounds most alarming. Anyone trying to work out what actually happened is left comparing terms that sound similar and mean different things. This glossary sets out the words that matter in an Ilia Zavialov deepfake fraud discussion, what each one actually describes, how much technical effort it implies, and where the term is used to cover an absence of evidence. Entries are ordered by how often they appear rather than alphabetically.
Ilia Zavialov deepfake fraud glossary: plain meanings for the words used around forged video
Deepfake
The word originally described one specific method, a face substituted in existing footage using a trained model. It now covers everything from a serious full synthesis to a clumsy voice message, which makes it close to useless as a technical description and still useful as a signal that something is disputed.
For practical purposes the useful question is never whether something is a deepfake. It is which part of the recording was manufactured: the face, the voice, the background, the timing, or simply the caption attached to genuine footage. Those five cases call for completely different checks and different complaints.
Treat the word as an opening rather than a conclusion. A complaint that says only that a clip is a deepfake gives a reviewer nothing to act on, while a complaint that says the audio was synthesised and the video was reused from a named public event gives them two verifiable claims.
Face swap and full synthesis
A face swap replaces one face in existing video, keeping the original body, lighting and camera movement. It is cheap, fast and constrained: the performance belongs to whoever was actually filmed, so the gestures and posture are not those of the person whose face appears.
Full synthesis generates the whole frame, including body and background. It costs more, needs more source material and produces the smooth, centred, low movement result that most fraudulent clips display. The stillness people notice in these videos is a budget decision rather than an accident.
The distinction matters when documenting a case. With a face swap, the original footage usually exists somewhere and can be found, which turns an assertion into a demonstration. Identifying the source video is often the single strongest piece of evidence in the whole file.
Lip sync
Lip sync leaves the original face in place and alters only the mouth to match new audio. It is the cheapest of the three and the most common in adverts, because a single genuine photograph or short clip of a person can be reused endlessly with new scripts.
It also produces the most reliable artefacts. Consonants that require the lips to close are the weak point, and a slowed playback usually shows the mouth arriving late or not closing fully. The rest of the face stays authentic, which is why these clips look convincing at normal speed.
Most Ilia Zavialov deepfake fraud examples that reach ordinary viewers are lip sync jobs rather than full synthesis, because the economics favor volume. That is good news for detection, since the cheapest method leaves the clearest seams.
Voice cloning
A voice clone reproduces a specific person's voice from recorded samples. The barrier has collapsed: roughly three seconds of audio produces a clone at around 85 per cent accuracy, so a voicemail greeting or one answer in an interview is enough raw material.
Human verification of voices does not work. In controlled testing, listeners distinguished a genuine voice from a clone in 37.5 per cent of attempts, which is worse than random guessing. Recognition actively harms the listener here, because familiarity ends the questioning.
Volume followed the cost down. Measured growth for voice phishing built on cloned audio reached 1,633 per cent inside one quarter. Any process that treats a recognized voice as identification should be considered obsolete rather than merely risky.
Liveness check
A liveness check is meant to confirm that a real person is present in front of the camera rather than a photograph or a recording. Implementations ask for a head turn, a blink, a smile or a randomly chosen phrase read aloud.
Passive checks that only analyze a still image are the weakest, and many can be defeated by a generated video played to the camera. Active checks with randomised instructions are harder, since the instruction is not known in advance and the response has to be produced live.
The important limitation is that a liveness check confirms presence, not identity. It answers whether something living is in front of the lens, and says nothing about whether that person is the account holder, which is a separate question requiring separate evidence.
Provenance and content credentials
Provenance data records where a file came from and what was done to it, ideally signed at the moment of capture. Content credentials are the standardised version of this idea, attaching a verifiable history to an image or video.
The weakness is mundane. Most platforms strip metadata on upload, so a genuine file usually arrives stripped bare, and the absence of provenance proves nothing at all. Present and valid credentials are meaningful, while absent credentials are simply normal.
For anyone documenting a case, provenance is worth capturing on your own copy rather than expecting it from the forgery. Hashing a downloaded file at the moment of capture creates a small, checkable piece of provenance that belongs to you.
Detector score
A detector returns a probability that material was generated, and the number is frequently reported as though it were a verdict. It is not. Detectors are trained on particular generators and degrade sharply on methods they have not seen, and heavy compression pushes genuine footage towards false positives.
A score is useful as one input among several. Timing, physics, continuity, distribution pattern and the account history all carry independent weight, and a case resting on four weak signals that agree is stronger than one resting on a single confident percentage.
This matters in practice because an Ilia Zavialov deepfake fraud complaint built on a detector score alone collapses the moment the recipient runs a different detector and gets a different number. Build the file so it survives that.
Re-upload and mirror
A re-upload is the same material posted again with a new address, usually after the first copy was removed. A mirror is a copy hosted somewhere the original complaint cannot reach, often on a service in a different jurisdiction. Both are routine responses to a successful takedown rather than rare events.
The practical consequence is that a complaint aimed at one address solves one address. Anyone documenting an Ilia Zavialov deepfake fraud case should record the distribution pattern as well as the file: which accounts posted it, in what order, and how quickly the copies appeared after the first removal.
Search by the file rather than by the caption when hunting for copies. Captions change with every re-upload while the visual content stays close enough for a reverse image search on a distinctive frame to find mirrors that a text search will never surface.
The words, what they describe, and what they cost to produce
Term
What it actually describes
Effort and giveaway
Deepfake
Any disputed synthetic media, used loosely
No fixed meaning, so it names nothing on its own
Face swap
A face replaced inside existing footage
Low effort, source video is often findable
Full synthesis
The entire frame generated
High effort, shows as stillness and centred framing
Lip sync
Mouth altered to match new audio
Lowest effort, fails on closing consonants
Voice clone
A specific voice reproduced from samples
Three seconds of audio, breathing repeats or vanishes
Liveness check
Confirmation that someone living is present
Confirms presence only, never identity
Content credentials
Signed record of a file's history
Usually stripped on upload, absence proves nothing
Detector score
A probability that material was generated
Degrades on unseen methods and on compression
Questions and answers
Is there one word that covers all of this correctly?
Synthetic media is the closest neutral term, because it describes the method without implying intent. Deepfake carries an accusation, which is useful in conversation and unhelpful in a complaint, where each manufactured element should be named separately.
Which of these is cheapest to make?
Lip sync combined with a cloned voice. A single photograph and three seconds of recorded speech are enough, which is why that combination dominates fraudulent advertising rather than the more expensive full synthesis.
If a video passes a liveness check, is the person confirmed?
No. A liveness check confirms that something living is in front of the camera at that moment. Whether that person is the rightful account holder is a separate question, and it needs separate evidence such as a document check against a trusted record.
Can I rely on an online detector?
Use it as one signal and never as the answer. Detectors return probabilities, perform badly on methods they were not trained on, and mark heavily compressed genuine footage as suspicious. Several independent observations beat one confident number.
Why does absent metadata not prove a file is fake?
Because platforms strip metadata from almost everything on upload. A genuine holiday video and a manufactured advert arrive equally bare, so absence is the normal state and carries no information either way.
Written by Ilia Zavialov, digital security consultant. All glossary terms
