Law enforcement
Authenticate video evidence and produce court-admissible forensic reports.
How the sector deploys itDeepVideoGuard
Know whether a video is real footage or an AI-generated fake. It runs many independent checks across the picture, the motion and the sound at once, then produces a single clear verdict together with an annotated report you can share, keep on file, or submit as evidence
Product Walkthrough
0:000:00
What it is
Deepfake video forensics
DeepVideoGuard helps organizations determine whether a video is authentic or shows signs of AI generated, altered, or manipulated content. It examines multiple signals across the video, including visual details, movement, facial behaviour, sound, and the consistency between what is seen and heard.
Instead of relying on a single clue, the platform brings together different layers of analysis to identify inconsistencies that may be difficult to notice through manual review alone. These findings are combined into a clear authenticity assessment, helping users understand whether the submitted footage appears real or requires further investigation.
Each analysis produces an easy to understand result supported by visual findings and a forensic report, giving investigators, security teams, media organizations, businesses, and other reviewers a practical way to verify video authenticity and document their findings.
A fake has to slip past all of them, not just one.
- Specialised checks run in parallel on every video
8
Specialised checks run in parallel on every video
- Vision-language reasoning, the heaviest weight in the verdict
30%
Vision-language reasoning, the heaviest weight in the verdict
- Annotated diagnostic video with frame-synced score overlays
1080p
Annotated diagnostic video with frame-synced score overlays
- Cryptographic file hash carried in the forensic report
SHA-256
Cryptographic file hash carried in the forensic report
Try it here
Run a video through the eight modules
Drop a video here
MP4, MOV, WebM or MKV up to 200 MB
The eight modules start as soon as a file lands · 1 credit
1 credit per analysis
Run the same eight modules the platform runs — on a video of your own, with every module’s score.
3 / 3Free analyses left
See credit termsPreview of the output format. The module scores and the verdict shown here are derived in the browser from the file’s digest, not measured by the analysis engine — the SHA-256 is the one figure computed from the real bytes. Book a walkthrough to have footage of your own run through the live ensemble.
How it works
From an uploaded video to an evidence-ready verdict
01
Upload
Securely submit the video you want to verify for authenticity
02
Prepare
The video is securely prepared for analysis and integrity verification
03
Analyse
Visual, motion and audio signals are examined together for signs of AI generation or manipulation
04
Assess
The findings are brought together to determine whether the video appears authentic or potentially synthetic
05
Report and video
Receive a clear video authenticity result with supporting visual findings and an evidence ready forensic report
Product exclusiveness
What sets DeepVideoGuard apart
Eight-module ensemble engine
Rather than trusting one detector, DeepVideoGuard runs eight specialised checks in parallel — from vision-language reasoning to motion tracking, frequency analysis and audio-visual sync — each hunting for a different tell-tale sign of manipulation. A fake has to slip past all of them, not just one.
Flagship capability · eight checks in parallel
Weighted verdict system
Every module's result is combined into one decision by weighting each check on its proven reliability — vision-language 30%, temporal 15%, frequency 14%, and so on — so no single module can swing the outcome on its own.
Vision-language 30% · temporal 15% · frequency 14%
Multimodal vision-language reasoning
A high-capacity vision-language model reviews key frames alongside the audio to catch what a pixel test misses — implausible facial expressions, broken scene logic and context that does not add up.
Key frames read alongside the audio
Annotated diagnostic video
Renders a side-by-side forensic video at 1080p with frame-synced score overlays and per-module gauges, so a reviewer can see exactly where — and why — the footage looks synthetic.
1080p side-by-side · per-module gauges
Evidence-ready PDF forensic report
Generates a shareable report containing the file's cryptographic hash (SHA-256), an executive verdict summary, the influence-weighting breakdown and each module's confidence — suitable for audit trails and legal review.
Audit-grade · SHA-256 file hash
Physics & corneal-reflection analysis
Checks whether the light reflected in the eyes follows real optics and lighting geometry, catching the impossible or asymmetric reflections that generated faces so often get wrong.
Eye optics and lighting geometry
Lip-sync & audio-visual alignment
Measures how closely spoken sounds line up with lip movements to expose dubbing, voice cloning and the subtle mouth mismatches that betray a swapped or synthesised speaker.
Spoken sound against lip movement
Frequency, texture & GAN-fingerprint analysis
Inspects faces in the frequency domain and searches for the micro-patterns left behind by generative models — GANs and diffusion — revealing artefacts that are invisible to the naked eye.
Frequency domain · GAN and diffusion traces
How it compares
Measured against a single-model deepfake detector and traditional manual video review
Detection approach
- Single-model detector
- Partly — Single model
- Manual video review
- No — Human eye
- DeepVideoGuard
- Yes — 8-module ensemble
Combines many signals into one verdict
- Single-model detector
- No — Single score
- Manual video review
- Partly — Subjective
- DeepVideoGuard
- Yes — Weighted consensus
Analyses audio together with video
- Single-model detector
- No — Often video-only
- Manual video review
- Partly — Manual
- DeepVideoGuard
- Yes — Lip-sync + VLM
Checks physical plausibility (eye optics, lighting)
- Single-model detector
- No — No
- Manual video review
- Partly — Expert only
- DeepVideoGuard
- Yes — Physics module
Shows where the manipulation is
- Single-model detector
- No — Score only
- Manual video review
- Partly — Written notes
- DeepVideoGuard
- Yes — Annotated video
Evidence-grade report
- Single-model detector
- No — No
- Manual video review
- Partly — Written statement
- DeepVideoGuard
- Yes — SHA-256 PDF
Detects GAN & diffusion fingerprints
- Single-model detector
- Partly — Varies
- Manual video review
- No — No
- DeepVideoGuard
- Yes — Frequency + residual
Scales for volume
- Single-model detector
- Partly — Varies
- Manual video review
- No — Manual throughput
- DeepVideoGuard
- Yes — Real-time job queue
The architecture
Eight checks in, one verdict out
No single module can swing the outcome on its own
What goes in
One video
- Vision-language30%
- Temporal motion15%
- Frequency (DCT)14%
- ViT artefacts12%
- Physics & optics10%
- Lip-sync8%
- Zero-shot6%
- GAN fingerprint5%
Weighted consensus · 100%
Each check is weighted on its proven reliability, so no single module can swing the outcome on its own.
Authentic or synthetic
- Forensic reportPDF · carries the SHA-256
- Annotated video1080p · per-module gauges
Use cases by sector
Use cases across sectors
Legal & courts
Establish whether submitted footage is genuine or synthetically generated.
How the sector deploys itMedia & journalism
Verify newsroom and user-submitted video before publication.
Government & elections
Counter video disinformation targeting officials and campaigns.
How the sector deploys itCorporate security
Investigate suspected deepfakes used in fraud and impersonation.
How the sector deploys itDeployment strategy
Upload a video, get a verdict and a report
Upload the video
Submit through the web app (with authenticated sign-in) or the API. The file is hashed with SHA-256 for integrity and placed on the job queue.
Distributed to GPU workers
A message broker hands the job to accelerated engine workers that run the eight detection modules, processing picture, motion and sound in parallel.
Live progress
Watch per-module progress update in real time while the analysis runs asynchronously in the background — no need to wait on a blank screen.
Weighted verdict
Module scores are combined by the weighted-consensus engine into a final authentic-or-synthetic verdict with an overall confidence figure.
Collect report & diagnostic video
Download the evidence-ready PDF and the annotated 1080p diagnostic video. Media is stored securely and delivered only to authenticated users.
Under the hood
What the platform is made of
The architecture, the processing, the security model, and what you get back.
- Architecture
- A decoupled set of services — a web front end, an API back end, a message broker for job queueing, accelerated engine workers and a relational store for state and evidence.
- Processing
- Accelerated workers run all eight modules. Jobs are queued and processed in the background with API rate limiting, so large or bursty workloads stay orderly.
- Security
- Zero-trust access: email OTP verification, bcrypt password hashing and stateless JWT sessions, with protection against object-reference tampering, HTTP security headers (HSTS, CSP, X-Frame-Options) and authenticated media delivery.
- Outputs
- A SHA-256-hashed PDF forensic report and an annotated H.264/AAC 1080p diagnostic video, both released only to authenticated users.
Frequently Asked Questions
It runs eight specialised checks in parallel across the picture, the motion and the sound, so a fake has to slip past all of them rather than just one. These include multimodal vision-language reasoning over key frames and audio; temporal / motion tracking and frequency-domain analysis; physics and corneal-reflection checks on eye optics and lighting geometry; lip-sync and audio-visual alignment to expose dubbing and voice cloning; and GAN and diffusion fingerprint detection for generator artefacts. Each module's result is combined into one decision by a weighted-consensus engine, so no single module can swing the outcome on its own.
Two outputs: an evidence-ready PDF forensic report and an annotated 1080p diagnostic video. The report contains the file's SHA-256 hash, an executive verdict summary, the influence-weighting breakdown and each module's confidence. The video shows side-by-side frame-synced score overlays and per-module gauges, so a reviewer can see exactly where — and why — the footage looks synthetic.
Yes. The report is designed to be audit-grade — the cryptographic hash establishes file integrity, and the per-module breakdown documents how the verdict was reached. It is intended to be shared, kept on file or submitted for audit trails and legal review, and is used by law enforcement and courts for exactly this purpose.
Submit through the authenticated web app or the API. The file is hashed on upload and placed on a job queue that distributes it to GPU workers running all eight modules. You watch per-module progress update in real time rather than waiting on a blank screen, and jobs are queued with API rate limiting so large or bursty workloads stay orderly.
Access is zero-trust: email OTP verification, strong password hashing and stateless session tokens, with protection against object-reference tampering and standard HTTP security headers. Both the report and the diagnostic video are stored securely and delivered only to authenticated users.
The rest of the line
More in SyntheticMediaGuard
Verify video evidence with 8-module ensemble intelligence
Schedule a technical demonstration to run your video footage through our Multimodal VLM and forensic analysis engine.