WatermarkDeepfakeDetector<T>
Models & Types · AiDotNet.Safety.Audio
Detects AI-generated audio by looking for the presence or absence of known watermark patterns (e.g., AudioSeal-style localized watermarks).
Inheritance: ModelBase<T, Vector<T>, Vector<T>> → SafetyModuleBase<T> → AudioSafetyModuleBase<T> → WatermarkDeepfakeDetector<T>
For Beginners
Some AI companies embed invisible “watermarks” in the audio they generate — like a secret signature proving it’s AI-made. This module looks for those signatures. If found, we know it’s AI-generated. If we also detect other AI artifacts but no watermark, it might be an attempt to hide AI origin.
How It Works
Many responsible AI audio generators embed watermarks into their output. This detector checks for characteristic patterns of known watermarking schemes: spectral energy in specific sub-bands, phase-based encoding, and spread-spectrum signatures. The absence of any camera/microphone artifacts combined with no watermark is also a signal.
References:
- AudioSeal: Localized watermarking for speech (Meta AI, 2024, arxiv:2401.17264)
- SoK: Systematization of watermarking across modalities (2024, arxiv:2411.18479)
- Only 38% of AI generators implement adequate watermarking (2025, arxiv:2503.18156)
Constructors
| Constructor | Summary |
|---|---|
WatermarkDeepfakeDetector(double threshold, int sampleRate) |
Initializes a new watermark-based deepfake detector. |
Properties
| Property | Summary |
|---|---|
string ModuleName { get; } |
Methods
| Method | Summary |
|---|---|
IReadOnlyList<SafetyFinding> EvaluateAudio(Vector<T> audioSamples, int sampleRate) |