Deepfake Detector
Definition
A machine-learning classifier trained to distinguish authentic recordings from AI-synthesised or face-swapped media. Outputs a probability score rather than a binary verdict. Published false-positive and false-negative rates are central to interpreting the score in a report.
- Output type
- Probability score, not a binary verdict
- Function
- Distinguishes authentic recordings from AI-synthesised media
- Key reporting metric
- Published false-positive and false-negative rates
- Field
- Media authenticity examination
- Validation need
- Requires testing on data matching the case's media type and quality
Common questions
Why should a deepfake detector's probability score never be reported as a simple yes or no finding?+
A raw score reflects the model's confidence given its training data and does not translate directly into certainty about the specific footage under examination. Reporting it as a bare binary conclusion hides the false-positive and false-negative rates that a court needs to weigh the finding properly.
What can undermine the reliability of a deepfake detector's output in a given case?+
Detectors are trained on particular generation methods and video qualities, so heavy compression, an unfamiliar generative technique, or content very different from the training distribution can produce unreliable scores, which is why published error rates must be tied to conditions resembling the actual evidence.
Related terms
- C2PA (Coalition for Content Provenance and Authenticity)
- An open technical standard that embeds cryptographically signed provenance assertions into media files at the point of capture or editing. A C2PA...
- Error Level Analysis (ELA)
- A technique that re-compresses a JPEG at a controlled quality setting and maps the pixel-level difference between the re-compressed and original images....
- False-Positive Rate
- The proportion of authentic media files that a detection tool incorrectly classifies as manipulated or synthetic. Deepfake detectors and ELA tools both...
- Hash Verification
- The process of computing a cryptographic hash (SHA-256 or equivalent) of the exhibit file and comparing it against a previously recorded value...
- Hedging Language
- Qualified phrasing that accurately conveys the degree of certainty a method supports. Examples: 'consistent with', 'the findings are indicative of', 'cannot be...