WebRTC media interception
Patches RTCPeerConnection before page scripts run to passively capture video and audio tracks. The original media is never modified, delayed, or degraded.
Open source project
Deepfake detection for video calls
TrueStream sits quietly in your browser and watches incoming video and audio for signs of manipulation. When something looks off, it tells you. When you need certainty, it lets you prove who you are through the Vinsium cryptographic identity protocol.
The problem
You cannot trust what you see
Real-time deepfakes can impersonate anyone on a video call. The person you are talking to might not be who they appear to be, and your eyes alone cannot tell the difference. TrueStream provides an early warning when something looks synthetic, and a path to definitive proof when AI confidence alone is not enough.
Features
What you get out of the box
Local ONNX inference
EfficientNet-B0 for face manipulation and RawNet2-lite for voice synthesis run in dedicated Web Workers via ONNX Runtime WASM. Under 5% CPU on modern hardware.
Five-level trust ladder
Trust is not binary. Scores map to Confident, Uncertain, Suspicious, and Likely Synthetic levels — plus a Verified level that only cryptographic proof can reach.
Vinsium cryptographic escalation
When AI confidence is not enough, challenge-response verification through the Vinsium protocol provides mathematical proof of identity. Optional — TrueStream works fully without it.
Platform adapters
Google Meet, Microsoft Teams, and Zoom Web are supported. Auto-detection based on hostname, with per-platform handling for DOM and media quirks.
OTA model updates
Models are not baked into the extension. A daily manifest check downloads new models with SHA-256 verification, allowing rapid response to new attack techniques.
Architecture
How it works
TrueStream intercepts WebRTC media tracks, extracts frames and audio at controlled rates, runs two quantized neural networks in Web Workers, and feeds scores through a composite scorer with asymmetric hysteresis to prevent alert flickering.
RTCPeerConnection (patched at document_start)
|
┌────┴────┐
| |
Video Audio
Track Track
| |
VideoTap AudioTap ──▶ AudioWorklet (audio thread)
1 fps 2s chunks accumulates PCM samples
224x224 16 kHz
| |
▼ ▼
Worker 1 Worker 2
EfficientNet-B0 RawNet2-lite
(ONNX WASM) (ONNX WASM)
5.8 MB INT8 1.4 MB INT8
| |
└────┬────┘
|
CompositeScorer (60% video / 40% audio)
|
TrustStateMachine (3 to degrade / 5 to recover)
|
┌────┴────┐
| |
Overlay Sidepanel
TrustRing Score timeline
(Shadow DOM) Verify buttonTrust model
The trust ladder
Trust is a spectrum, not a binary. TrueStream maps composite scores to five levels, plus a sixth that only cryptographic proof can reach. The design philosophy: certainty is reserved for mathematics, not statistics.
Verified
Cryptographically proven via Vinsium. Immune to AI-based degradation.
Confident
Score above 0.75. Consistent authentic signals from both video and audio.
Uncertain
Score 0.40 to 0.74. Middle ground. Vinsium verification suggested.
Suspicious
Score 0.15 to 0.39. Some manipulation indicators detected.
Likely Synthetic
Score below 0.15. Strong evidence of manipulated media.
Privacy
Everything stays local
No media leaves your device
All inference runs locally in the browser. No video frames, no audio samples, no screenshots are ever sent to any server.
No telemetry
Zero analytics, zero usage tracking. The only network requests are anonymous model manifest checks and optional Vinsium challenge-response (nonces and signatures only).
License
Open source
TrueStream is licensed under the Apache License 2.0. Free to use, modify, and distribute for any purpose.
Requirements
Chrome or Edge 110+
- Build from source —
npm run build - Development —
npm run devwith hot reload - Platforms — Google Meet, Microsoft Teams, Zoom Web
Trust, but verify
TrueStream treats the arms race between detection and generation realistically. Rather than claiming perfect accuracy, it embraces uncertainty through the trust ladder and provides a mathematical escape hatch via Vinsium when AI alone is not enough.