Skip to content
ExiphoreRequest a demo
Case studies

Technical case study · Law enforcement

A circulated video of a serving officer

A 47-second clip of a senior Delhi Police officer announcing his resignation spread across messaging apps. This is the complete examination: what twenty signals measured, what the finding was, and the three things it could not establish.

A Delhi Police officer on duty beside orange police barricades in New Delhi
Exhibit
Video · 47.3 s · 360 × 444 · 167 kbps
Faces detected
1 · 95 px wide
Examination profile
DEEP · 27.9 s wall clock
Manipulation index
0.74 · interval 0.59 – 0.88
Examiner confidence
57% · capped by exhibit quality
Signals used
13 of 14 applicable

What arrived

A video circulating on messaging apps showed a serving senior officer of the Delhi Police appearing to announce his resignation and to make allegations about the conduct of an investigation. The clip ran 47.3 seconds at 360 × 444 and 167 kbps — a file that had been forwarded, re-encoded and compressed repeatedly before anyone thought to examine it.

That degradation matters more than anything else in this case, and the report says so on its first page. A denial is not a finding. What the force needed was something that would still hold weeks later, when it was challenged.

Intake and sealing

The file was hashed with SHA-256 on arrival and written to a content-addressed, write-once locker. From that point the hash is the identity of the exhibit: the seal is re-verified before every examination, and a mismatch refuses the run rather than reporting a result against altered bytes.

Every action from intake onward was written to a hash-chained chain of custody and a separately hash-chained audit log. At the close of this examination the chain stood at 58 entries, each carrying the hash of its predecessor.

What the examination measured

Twenty signals ran across eight independent families of forensic evidence. Each family measures something different about the exhibit, and — this is the point of running eight — a generator that defeats one is not thereby defeating the others. Three families carried the finding:

  • Physiological, 41.3% of the finding — blood-volume pulse recovered from four facial regions disagreed between regions: 98, 61, 48 and 100 bpm. A real face perfused by one heart does not do that.
  • Frequency, 29.1% — the high-frequency tail of the power spectrum departed from the natural-image power law, the signature of a generator's upsampling stage.
  • Temporal, 27.7% — learned lip-sync analysis was inconsistent with genuine capture across the sequence.
  • Metadata contributed 1.9%: the acquisition structure had been destroyed by transcoding, which is normal for redistributed media and therefore weighted low.
  • Artifact and Learned-model families returned 0.0% — they found nothing. That is recorded as finding nothing, not as evidence of authenticity.
  • Provenance was excluded entirely: no C2PA content credential or signature was recoverable, which is uninformative for a re-encoded file and so carries no weight in either direction.

The finding

Manipulation indicated, at a manipulation index of 0.74 on a 0–1 scale, with a confidence interval of 0.59 to 0.88. The interval is the honest width of the result: a wide band means the signals disagree, or the exhibit is too degraded to narrow it.

Examiner confidence was 57% — and the report states why. The quality ceiling imposed by the exhibit was also 57%: a single face, 95 pixels wide, in a heavily compressed frame. No examination of this file could have returned a higher confidence, whatever it found. A platform that reported 95% on this exhibit would be lying to you.

Thirteen of fourteen applicable signals were used. The adjudication was fixed before any narrative was generated, and the examination closed pending a named examiner's signature — an automated result is not released as an expert opinion.

What linked it to earlier work

The exhibit was matched against the registry of everything previously examined on the instance. It linked at 100% on perceptual hash, 100% on generator fingerprint, and matched on face embedding to a record already held.

Perceptual hash and generator fingerprint are strong technical links: the same source file, and the same generator. Face embedding is an investigative lead and is reported as one — it says two exhibits show the same face, not that a named person made either of them.

What it could not establish

Three things, all of them in the report rather than in a footnote:

  • Attribution. The examination establishes how the media was made, not who made it or why. Nothing in it names an author.
  • Authenticity of the unexamined. A clean result on the families that found nothing does not mean the exhibit is genuine. Absence of evidence is not evidence of absence, and the report refuses that inference explicitly.
  • A confidence above the exhibit's own ceiling. 57% was the maximum this file could support. The remedy is a better copy, not a better score.

What the force received

A sealed examination report carrying the original's hash, the method, every signal with its contribution and its limitations, the index with its interval, the quality ceiling, the cross-case links, the complete chain of custody, and space for a named examiner's adjudication and signature.

The examination ran in 27.9 seconds on the force's own infrastructure. No part of the exhibit left the network.

What this study is not

  • The imagery in this study is mosaicked in the file itself. The clip is a defamatory synthetic of an identifiable public servant, and demonstrating that Exiphore catches it does not require republishing the allegation.
  • Figures are from one examination of one exhibit. They describe this case and are not a general accuracy claim; our measured performance is published separately, with its sample and its limitations.