Deepfake detection for investigators.
Forensiqa combines AI with dynamical systems to model how real faces, voices and scenes behave, and flags media that breaks those patterns. It learns from far less data than conventional detectors, and adapts to new generators with a few of your own examples.
To whom it may concern, I am writing about the incident on the evening of March 2. I was at the location until around 11 pm and left before anything happened.
It is important to note that I have always maintained a respectful and professional relationship with all parties involved, and I remain committed to transparency throughout this process.A1 I did not see the other driver after leaving.
My phone records will show I was home. I want to emphasize that I value honesty, I value accountability, and I value the trust that has been placed in me.A2 Please contact me if you need anything else.
Interactive preview. Product screens show sample data.
Detection that understands how real media behaves.
Most detectors memorize what yesterday's fakes looked like, so they need huge datasets and fall behind each new generator. Forensiqa models what real media looks like instead.
Built on dynamical systems
Our models learn the rules that govern real signals over time: how a face moves, how a voice is produced, how light falls on a scene.
Learns from little data
Because the structure is built into the model, Forensiqa reaches high accuracy from a small fraction of the training data other approaches need.
Fine-tune on your data
Train a private model on your own cases, or on a new generator you've just encountered, with a handful of labeled examples.
Every finding explained
Each result points to the moment, region or passage where the media stopped behaving like the real thing.
Real media follows patterns. Fakes break them.
Faces move, voices are produced and light falls according to physical rules. Forensiqa learns those rules as dynamical systems, and measures how far a piece of media drifts from them.
We model how real signals evolve, then measure the deviation.
Forensiqa learns the dynamics of authentic media: the trajectories facial landmarks follow, the rhythm of breathing and speech, the way texture and lighting hold together across a scene. A generated or edited file can look right frame by frame and still move in ways real ones never do.
- Models the underlying behavior of real signals, not the artifacts of one generator
- Prediction error shows exactly where and when media departs from real dynamics
- Holds up better as new generation tools appear
Train it on your own cases, in minutes.
When you meet a generator Forensiqa hasn't seen, or media specific to your work, upload a few labeled examples. Forensiqa trains a private version of the model, tests it on examples it hasn't seen, and locks the version so every result can be reproduced.
- Start from dozens of labeled examples
- Automatic holdout testing with error rates before you deploy
- Private to your organization, versioned and locked like every model
See why every file was flagged.
Because Forensiqa measures deviation from real behavior, every finding comes with a reason: the segment of audio, the region of a frame or the passage of text where the pattern broke.
- Findings located on the timeline, in the frame or in the text
- Plain-language explanation for each finding
- Inconclusive results labeled as inconclusive
Wherever a fake can change the outcome.
Teams use Forensiqa to check media before they act on it, from a single viral clip to thousands of files from a seized device.
Digital evidence
Authenticate bodycam, CCTV, phone media and documents before they shape an investigation.
Fraud and impersonation
Check voice notes, calls and video meetings used to impersonate executives, relatives or officials.
Identity verification
Screen selfies, ID photos and liveness videos for face swaps and generated identities.
Disinformation and OSINT
Verify viral clips and images by URL before they're reported, shared or acted on.
Image-based abuse
Help investigators and victim-support teams identify non-consensual synthetic intimate imagery quickly.
Insurance claims
Catch generated or edited photos of damage, receipts and documents submitted with claims.
Results that hold up when the case goes to court.
Every file is hashed at upload and every action is logged. Each analysis is locked to its model version, including your custom models, so anyone can rerun it and get the same answer. Reports include the method, the error rates and the full custody log.
| Time | Action | By | Hash |
|---|---|---|---|
| Oct 14 09:12:07 | Received upload | an-112 | 9f2c…e41a |
| Oct 14 09:14:52 | Viewed | an-112 | match |
| Oct 14 10:20:31 | Analysis R-2214 | an-112 | match |
| Oct 16 11:02:10 | Report exported | ex-031 | match |
| Oct 21 14:37:55 | Rerun R-2290 · identical | dx-008 | match |
Use it in the browser, or build it into your tools.
Analysts work in the web app. Engineering teams connect Forensiqa to case management, KYC flows and moderation queues, and train custom models, through the API.
Web app
Upload files or paste URLs, review findings, manage cases, train custom models and export reports.
REST API
Send files, get structured results and webhooks, and start fine-tuning jobs programmatically.
Be one of the first teams to use Forensiqa.
We're working with a small group of investigation teams ahead of launch. Request early access, or book a demo to talk through your use case.
Thanks.
Preview build: this form isn't connected yet, so nothing was sent.