Deepfake Forensics and Media Authentication

What a media authentication covers

The most reliable evidence about a file is usually not inside it. Where it came from, and what happened to it on the way, decides more cases than any artifact in the pixels.

Provenance and acquisition path

The examination begins by establishing the earliest available instance of the media and acquiring it forensically — from the capturing device, the originating account, or the platform that received it — rather than accepting a copy that has been forwarded, re-encoded, and stripped along the way. That acquisition is hashed and recorded, and the difference between the acquired original and the copy in the record is itself a finding.

• Container and codec structure — the internal layout of an MP4 or MOV file, the brand and encoder strings it carries, quantisation tables in a JPEG, and whether all of that is internally consistent with the device the file is said to have come from. Editing tools and generative pipelines leave structural signatures that differ from camera-native output.
• Embedded metadata — capture timestamps, device make and model, lens and exposure data, geolocation, and edit history, tested against each other and against the surrounding record rather than accepted at face value.
• C2PA content credentials — cryptographically signed capture and edit manifests, increasingly embedded by cameras, phones, and editing software. A valid manifest is strong affirmative evidence; its absence proves nothing, because most media in litigation predates or bypasses the standard.
• Corroborating records — call and message logs, cloud and camera-roll artifacts, backup history, surrounding communications, and the account activity of whoever distributed the file.

Signal-level analysis

Where provenance is inconclusive, the examination turns to the media itself. For video and images that means compression history — evidence of re-encoding and of regions quantised differently from their surroundings — sensor noise consistency across the frame, lighting and reflection geometry, frame-rate and interpolation artifacts, and alignment between lip movement and audio. For audio it means spectral examination for splices and discontinuities, consistency of the noise floor and room acoustics across the recording, codec and encoding history, and the spectral fingerprints left by voice-synthesis models. Where a recording captured mains hum, electrical network frequency analysis can sometimes place it in time.

Detectors are inputs, not conclusions

Automated detection tools have a role in triage and none in the conclusion. Their error rates move as generative models change, the datasets they were validated against rarely resemble the evidence in a given matter, and a single confidence score offers a court nothing it can test. Every tool used is named and versioned in the report, and what it can and cannot support is stated alongside its output. That is not caution for its own sake — it is what Rule 702 requires of the proponent.

Where this comes up

• Recordings offered as admissions, threats, or harassment, where the speaker denies ever saying it
• Fraud and financial matters where a voice on a call authorised a transfer
• Employment and internal investigations turning on surveillance or bodycam footage
• Family and protective-order proceedings where a single recording carries the allegation
• Defensive engagements answering a deepfake allegation made against genuine evidence

How the engagement runs

1. 01Conflicts and scope. Parties run for conflicts; the scoping call establishes what is being contested, what media exists, where the earliest copy might be, and the deadline.
2. 02Preservation. Devices and accounts holding earlier copies are preserved before anything else, along with the messaging or platform records that carry the distribution path. Camera roll and cloud photo libraries have their own deletion timers.
3. 03Acquisition. The earliest available instance is acquired forensically and hashed, together with the device or account artifacts around it. Every copy examined is logged with its source and hash so that the versions cannot be confused later.
4. 04Provenance examination. Container structure, metadata, content credentials, and distribution path, tested for internal consistency and against the claimed source device.
5. 05Signal examination. Compression, sensor, acoustic, and temporal analysis as the media and the allegation require, with each method and tool recorded.
6. 06Calibrated conclusion and report. Authentic to the claimed source, manipulated, or indeterminate — with the basis for the classification and the limits of the available material stated explicitly.

Authentication and the evidence rules

Authentication is a low threshold, and it is not the same question as reliability. A recording can clear Rule 901 and still be demolished on weight, and the forensic work usually matters more at the second stage than the first. Where the exhibit is a copy of a digital file, Rules 902(13) and 902(14) permit self-authentication on a qualified person’s certification with advance notice — which again depends on hashing the file at the moment it is acquired.

Where the rulemaking on AI and deepfakes actually stands

Two proposals have been before the Advisory Committee on Evidence Rules. As of the date of this page neither is in force, and neither has an effective date.

• Proposed Rule 707would require evidence produced by artificial intelligence and offered without a testifying expert to satisfy reliability requirements equivalent to those Rule 702 imposes on expert testimony. It was published for public comment in August 2025, and the comment period closed on February 16, 2026. At its meeting on May 7, 2026 the Committee did “not recommend action on the proposed Rule 707 at this time,” revising it instead and agreeing that the revised version “would require re-publication were it to go forward” — while deciding not to release it for a new period of public comment for now.
• A proposed Rule 901(c), directed squarely at deepfake challenges, has never been published for public comment at all. The working draft would require an opponent to present evidence sufficient to support a finding of fabrication before the court inquires at all, and would then require the proponent to demonstrate that the item is “more likely than not authentic” — a higher showing than the prima facie standard Rule 901(a) ordinarily applies. At the same meeting the Committee concluded that such an amendment “is not warranted” for now, noting how few courts have yet had to address the question, and is holding the draft in abeyance so that a rule would be ready without delay if that changes.

Both were referred for further study, with input from technology specialists, at the Committee’s next meeting. Neither changes the Rule 901 and 902 framework that governs a contested recording today. Confirm the current state of the rules, and any standing order in your forum, before relying on it. Source: Report of the Advisory Committee on Evidence Rules (May 17, 2026), in the Committee on Rules of Practice and Procedure agenda book for June 3–4, 2026, at 525, 534–37.

What you receive

Disclosure timing, deposition, and Rule 702 practice are covered on the expert witness testimony page.

Questions counsel ask

Related reading

• Authenticating deepfake audio and video

  Why provenance comes before pixels, what C2PA content credentials actually prove, and why a detector score will not survive Daubert.

• Daubert challenges to digital evidence

  Rule 702 as amended in December 2023, and how an overstated authentication opinion gets excluded.

• Computer forensics

  Where the earliest copy of a contested file usually lives, and how it is acquired without altering it.

• Mobile forensics and chain of custody

  Most contested recordings were captured on a phone. This is the custody discipline that keeps the original usable.