AudioWatermarkerBase<T>

Base Classes · AiDotNet.Safety.Watermarking

Abstract base class for keyed audio watermarkers that mark the block-DCT spectrum of the audio with a pattern derived from a secret key, and detect it by correlation.

Inheritance: ModelBase<T, Vector<T>, Vector<T>> → SafetyModuleBase<T> → AudioSafetyModuleBase<T> → AudioWatermarkerBase<T>

Implements: IAudioWatermarker<T>, IKeyedWatermarker

Known derived types: AudioSealWatermarker<T>, AudioWatermarker<T>, SpectralAudioWatermarker<T>

For Beginners

This base class provides common code for all audio watermarkers. The key works like a password: the watermark is a faint pattern spread over thousands of frequency values, too weak to hear in any one place, which adds up to strong evidence for whoever knows the exact pattern.

How It Works

The audio is cut into non-overlapping frames of about 32 ms (the smallest power of two of at least 0.032 x sample rate samples: 512 at 16 kHz, 2048 at 44.1 or 48 kHz). Each frame is transformed with the orthonormal DCT-II, and the coefficients inside the watermark band (BandLowHz to BandHighHz, clipped to the Nyquist frequency) are modified by the derived watermarker under the key’s +/-1 pattern for that frame (SignPattern, one stream per frame). The inverse DCT gives the marked frame back; samples after the last full frame are left as they are.

Detection computes, per frame, key-independent weights w_k from the received band coefficients (DetectionWeights) and the correlation z = sum(w_k x_k) / sqrt(sum(w_k^2)) over all frames. Because the weights do not depend on the key and the pattern bits are fair and independent, z is approximately standard normal for audio without this key’s mark, whatever the audio is. The confidence is 0 up to z = 3 (false-positive rate 0.135%) and 1 from z = 6; EvaluateAudio reports from ReportThreshold.

EmbedWatermark checks the result: if the marked audio does not reach z = 6 (too short, silent, or too low a strength), it throws instead of returning audio that would not be recognised. Frames are aligned to the first sample, so the mark survives noise, re-quantization and gain changes, but not cropping or time shifts that move the frame grid.

Properties

Property Summary
bool UsesPublicKey { get; }

Methods

Method Summary
IReadOnlyList<SafetyFinding> EvaluateAudio(Vector<T> audioSamples, int sampleRate)
Vector<T> EmbedWatermark(Vector<T> audioSamples, int sampleRate)
WatermarkResult Detect(Vector<T> audioSamples, int sampleRate) Detects this watermarker’s keyed watermark and reports the correlation test behind the confidence.
double DetectWatermark(Vector<T> audioSamples, int sampleRate)