Platform
Deepfake detection that leads with provenance
Deepfakes have moved from novelty to business-process risk — a forged invoice, a cloned voice on a call, a synthetic face in a verification flow. SAF3AI checks what can be proven before it asks a model to guess, across image, audio and video — and tells you plainly what it does not know.
Capabilities
Key Features
Provenance before prediction
C2PA credentials, EXIF and perceptual hashing run first, deterministically, on every item. Where provenance answers the question, no model is consulted at all — a signed credential is evidence, and a model score is an opinion.
Image, audio and video
All three are live. Images run a frozen-encoder probe alongside independent vision models; audio runs voice anti-spoofing; video combines frame-level and audio-track analysis with temporal consistency.
An ensemble, not a model
Two independent architectures score each image and the fusion stage measures whether they agree. Disagreement lowers confidence rather than being averaged away, which is what makes the confidence figure mean something.
Calibrated confidence bands
One verdict per item with a confidence band and the factors behind it, so a borderline result presents as borderline instead of as a false precision your analyst has to unpick.
Into the same incident pipeline
Media verdicts feed the Detection Fabric as another tier, keyed to the same entities as every other signal — so a synthetic asset in a workflow correlates with who used it and what it reached.
Three ways to call it
A command-line tool for one-off checks, an HTTP endpoint for your own pipelines, and a fabric worker that consumes media continuously. Same engine, same verdict shape, all three.
The pipeline
Cheap and certain before expensive and probabilistic
Running provenance first is not only faster — it means the confident answers come from evidence, and the model is reserved for the cases where no evidence exists.
Provenance
Deterministic · every item
C2PA content credentials, EXIF metadata and perceptual hashing against known assets. Cheap, explainable, and not a guess. Most items that can be resolved are resolved here.
ProducesSigned, altered, unknown
Detection ensemble
ML · only where provenance is silent
A router sends only unresolved media to the models. Images run a frozen CLIP encoder with a trained linear head plus two independent vision transformers; audio runs a WavLM-based anti-spoof probe; video combines both with temporal analysis.
ProducesPer-model scores and agreement
Fusion & calibration
Weighted, agreement-aware
Scores are calibrated and fused with the weight of each detector and the agreement between them, producing one confidence band rather than a raw number with no error bars.
ProducesOne verdict, with its factors
What we will not claim
Why provenance leads
Published deepfake detectors tend to look excellent on the dataset they were trained against and considerably worse on media from a generator they have never seen. That gap is well documented, it is not specific to any one vendor, and a detector that scores highly in-distribution can perform near chance on an unfamiliar source.
We test ours cross-source — trained on one dataset, evaluated on a completely different one — and we treat the worst number as the real number. That is why the architecture leads with provenance rather than with a model, why two independent architectures have to agree before confidence rises, and why an uncertain result is returned as uncertain.
Any vendor quoting a single headline accuracy figure for deepfake detection is quoting an in-distribution number. Ask what it scores on a generator it has not seen.
What this means in practice
- Signed media is verified, not guessed at.
- Altered media with intact provenance is caught deterministically.
- Unsigned media gets an ensemble score with an honest confidence band.
- Low-confidence results escalate to a person rather than auto-resolving.
Get started with Deepfake Detection
See how Saf3AI can help secure your AI agents.