Deepfake and Synthetic Media Forensics

What it is and why it matters

A deepfake is media produced or altered by a learned model. The common forms are a face swapped onto another body in video, a mouth re-animated to match new speech, a voice cloned from a few seconds of samples and images or whole clips generated from a text prompt. The tools are free or cheap, and the output has gone from obvious to convincing in a few years. The harms are already familiar in India: morphed images used for extortion, cloned voices used in fraud calls and fabricated clips of public figures.

The forensic question is not simply "is this fake". It is what evidence supports the claim that this media was captured by a camera or microphone as it appears, and what evidence supports the claim that it was generated or altered. Detection software answers with a score, which is a probability under the conditions the software was trained on, not a finding. Generators change every few months and a detector trained on last year's output can fail on this year's, in either direction. Compression, resizing and screen recording wipe out many of the traces detectors look for.

So the work rests on several legs. A trained examiner looks for the artefacts models still leave. Software adds a measured, documented opinion. Provenance data, where it exists, states how a file was made. And context, meaning where the file first appeared, who posted it and whether the events it shows can be checked elsewhere, often decides the matter faster than any pixel. The report says which legs the conclusion stands on and how firmly.

Types of digital evidence

Visual Generation Artefacts
Blending seams at the edge of a swapped face, mismatched lighting between face and neck, inconsistent reflections in eyes and glasses, oddly formed teeth, ears, hands and jewellery, garbled text and backgrounds that do not hold their shape.
Temporal Inconsistencies
Flicker between frames, a face that slides against the head when it turns, lips that do not match the sounds and movement that does not obey the physics of the scene.
Audio Cloning Traces
Flat or repeated prosody, unnatural breathing and pauses, spectral patterns left by the vocoder and a bandwidth or noise profile that does not match the claimed call or recording.
Statistical and Frequency Traces
Patterns in the pixel statistics and the frequency domain that some generators leave, measurable by software but fragile under compression.
Container and Metadata Signs
Software names, missing camera fields, encoder settings and structures that do not match any camera or phone, read with the caution that real files lose this data too.
Provenance Data
Content Credentials under the C2PA specification and the invisible watermarks some generators embed, which can describe or reveal a file's origin when present and intact.
Context and Source Trail
The earliest appearance of the file, the account that posted it, the platform's processing and whether the depicted event is corroborated by other records.
Reference Material
Genuine recordings of the person depicted, used to compare voice, face shape, mannerisms and the way they move.

Devices and data sources

Social media posts and the platform records behind them, obtained under lawful process
Messaging app forwards, usually several generations from the original
Recordings of calls and video calls in fraud cases
Images and clips sent in extortion and harassment complaints
Seized devices holding generation apps, prompts, model files and the source photos used to train a face
Accounts with online generation services, reached only with authority
Media produced by news organisations, public bodies and companies that is challenged as fake
Genuine reference recordings of the person depicted

How an examination runs

The order matters. Each step protects the one after it.

01
Fix the Question and Get the Earliest Copy
Agree what is in dispute. Obtain the earliest generation of the file available, with the records of where it came from, and hash it. A screen recording of a video call is a poor basis and should be stated as such.
02
Trace the Provenance
Check for Content Credentials and watermarks, find the earliest appearance online, read the posting account's history and note each platform the file passed through and what each one does to files.
03
Check the Context
Ask whether the event depicted could have happened: whether the person was where the clip suggests, whether other recordings exist and whether the content fits known facts. Context often settles what pixels cannot.
04
Survey the File
Read container, metadata and encoding details and compare them with what the claimed device or source produces.
05
Examine by Eye and Ear
Go through the material frame by frame and segment by segment, looking and listening for the artefacts listed above, with reference material of the real person at hand.
06
Run Detectors as One Input
Apply detection software, recording its name, version, training date, threshold and the validation evidence for media of this type and quality. Treat the score as an observation to be weighed, not a result to be reported alone.
07
Weigh and Conclude
Combine the provenance, context, structural, human and software findings and state the strength of support for each explanation. Where they conflict or the file is too degraded, say that no conclusion can be reached.
08
Report and Date It
Write the report with every limit stated, and note that detector results reflect the generators known at the time of examination and may need revisiting.

Tools used in this discipline

Independent profiles: what each tool does, where it stops and what you need to know before using it.

Browse the full tools directory

Terminology

Deepfake: Media in which a person's face, body or voice has been replaced or synthesised using a learned model. The word is also used loosely for any synthetic media.
Generative Adversarial Network: A pair of models trained against each other, one generating and one judging, used in earlier face generation. Its outputs carry characteristic frequency patterns that some detectors learned to find.
Diffusion Model: The approach behind most current image and video generators, which builds a picture from noise in many steps. Its traces differ from those of earlier models, which is one reason detectors age.
Face Swap and Reenactment: Face swap replaces the face in a video with another person's. Reenactment keeps the face and drives its expression and lip movement from another performance or from audio.
Voice Cloning: Generating speech in a target voice from a sample, sometimes only seconds long. Used in fraud calls and in fabricated recordings.
Detector Score: The probability-like number a detector produces. It depends on the training data, the threshold chosen and the quality of the input, and it is not a measure of truth.
Generalisation: A detector's ability to work on generators and conditions it was not trained on. Poor generalisation is the main reason scores drift and why validation on current material matters.
Content Credentials: Signed provenance data defined by the C2PA specification, recording how a file was made and edited. A verified manifest is strong evidence of history. Its absence says nothing either way.
Watermark: A signal hidden in generated media by the generator so its origin can be checked later. Only some generators embed one and processing can weaken it.
Liar's Dividend: The gain a person makes by dismissing genuine evidence as a deepfake. It is why examiners must be as careful in confirming real media as in exposing fakes.

Challenges and limits

What makes it hard

Detectors trained on one set of generators fail on others, so a score carries no meaning without evidence that the detector was validated on media of this kind
Compression, resizing and screen recording destroy the fine traces detectors depend on, and most evidence arrives in that state
Real low-quality footage can score as fake and good fakes can score as real, so a single score in either direction must not decide a case
Generators improve every few months and classic signs, such as odd blinking or six-fingered hands, disappear from newer output
Training data for detectors is unevenly spread across skin tones, lighting and languages, and performance varies with them
Provenance data is missing from almost all genuine media today, so its absence cannot be held against a file
Explaining a machine learning result to a court needs the method, its validation and its error rates, which vendors do not always disclose

Keeping the evidence intact

Obtain and hash the earliest available generation and record every platform it passed through
Record the detector's name, version, model date, threshold and validation evidence alongside its score
Keep the human examination notes separate from the software output so each can be assessed on its own
State the strength of support for each explanation rather than a verdict, and record where no conclusion was possible
Date the examination and note that findings depend on the generators and detectors known at the time

Standards and legal considerations

SWGDE Best Practices for Image Authentication
SWGDE's Best Practices for Image Authentication and Best Practices for Digital Video Authentication frame synthetic media as an authentication question, with structure, content and context examined together.
Read the guide
C2PA Technical Specification
The Content Credentials specification from the Coalition for Content Provenance and Authenticity. It defines the signed manifests that record how media was created and edited, the main provenance mechanism now being adopted.
NIST Open Media Forensics Challenge
NIST's evaluation series for manipulation detection in images and video, including generated content. Its datasets and results show how detector performance is measured and how far it varies.
ISO/IEC 27042:2015
The guideline on analysis and interpretation, which fits a field where the observation is a score and the interpretation is the examiner's.
Read the guide
Information Technology Act, 2000
India's law on offences such as cheating by personation using a computer resource and violation of privacy through images, under which many deepfake harms are investigated. Educational context only, not legal advice.
Read the guide

In practice

A Cloned Voice Call
A finance officer receives a call in the voice of a director asking for an urgent transfer and records part of it. The recording is narrow-band and short. The examiner finds flat prosody and no breath sounds, compares it with genuine recordings of the director and reports moderate support for a synthetic voice, while explaining that the call quality limits the finding. The call records and the account used matter as much as the audio.
A Morphed Image in an Extortion Complaint
A complainant receives an image placing their face on another body. The face region shows a blending boundary, its noise level differs from the rest of the image and the lighting on the face does not match the scene. The examiner reports strong support for a composite and describes the regions involved, without attempting to identify the generator.
A Genuine Clip Called a Deepfake
A party claims a CCTV clip of them is fabricated. The native export from the recorder is obtained, the structure matches the recorder's encoder, the system log shows continuous recording and the audio and video agree. The examiner reports no indications of synthesis or editing and explains the basis, which is a stronger position than any detector score.

This page is educational material, not legal advice. Laws and standards change, commence on different dates and are interpreted differently across jurisdictions. Always read the official text linked on this page and take advice from a qualified professional for a real matter.

Learn it at DFRTI

Free pages explain. These courses teach it on real devices with a mentor.

VideoSurveillance & Image / Video ForensicsDeepfake / AI ForensicsExamine face swaps and AI-made media, check C2PA provenance and know where detectors failAdvanced

AdvancedAdvanced & Specialised ForensicsAI / Machine Learning ForensicsUse machine learning to sort large evidence sets and investigate AI systemsAdvanced

AdvancedAdvanced & Specialised ForensicsOSINT / Open Source IntelligenceFind, verify and record public online information in a form a court can checkIntermediate

LegalFoundation & Legal CoursesLegal & Court ProceduresIndian law on electronic evidence: BSA Section 63, BNSS search and seizure and the IT Act.BeginnerDCFA

Careers that use it

Questions people ask

Can software tell whether a video is a deepfake?

Software can give a score that is useful on media similar to what it was trained on. It cannot settle the question alone, it is wrong on both real and fake material at times and its performance changes as generators change. It is one input to an examination.

Is a 90 per cent score proof?

No. The number depends on the training data, the threshold and the input quality. A 90 per cent score on a compressed WhatsApp forward from a detector never tested on such files is not evidence of anything. The report explains what the score means for this file.

Will Content Credentials solve the problem?

They help where they exist. A verified manifest records how a file was made, which is strong evidence. Most files today carry no manifest, and a file without one is not thereby fake. Watermarks have the same limit.

What should I do when I receive a suspected deepfake?

Keep the earliest copy you have and do not re-share it. Note where and when it first appeared and who sent it, then ask for an examination that covers provenance, context and content together. Report to the police where there is harm.

Can you tell which tool made a deepfake?

Rarely with confidence. Some generators leave distinctive traces or watermarks, but these are fragile and many tools share the same underlying models. The usual finding is whether the media is synthetic or altered, not which product produced it.