detector_benchmark.watermark.synth_id.detector

Classes

SynthIDDetector	Base class for SynthID watermark detectors.
MeanDetector	Base class for SynthID watermark detectors.
WeightedMeanDetector	Base class for SynthID watermark detectors.

Functions

get_detector(detector_name, logits_processor)

Module Contents

class detector_benchmark.watermark.synth_id.detector.SynthIDDetector

Bases: abc.ABC

Base class for SynthID watermark detectors.

This class defines the interface that all SynthID watermark detectors must implement. Subclasses should override the detect() method to implement specific detection algorithms.

abstract detect(g_values: numpy.ndarray, mask: numpy.ndarray) → numpy.ndarray

Detect watermark presence in the given g-values.

Args:

g_values: Array of shape [batch_size, seq_len, watermarking_depth] containing

the g-values computed from the text.

mask: Binary array of shape [batch_size, seq_len] indicating which g-values

should be used in detection. g-values with mask value 0 are discarded.

Returns:

Array of shape [batch_size] containing detection scores, where higher values indicate stronger evidence of watermarking.

Raises:

NotImplementedError: This is an abstract method that must be implemented by subclasses.

class detector_benchmark.watermark.synth_id.detector.MeanDetector

Bases: SynthIDDetector

Base class for SynthID watermark detectors.

This class defines the interface that all SynthID watermark detectors must implement. Subclasses should override the detect() method to implement specific detection algorithms.

detect(g_values, mask)

Args:

g_values: shape [batch_size, seq_len, watermarking_depth] mask: shape [batch_size, seq_len]

Returns:

scores: shape [batch_size]

class detector_benchmark.watermark.synth_id.detector.WeightedMeanDetector

Bases: SynthIDDetector

Base class for SynthID watermark detectors.

This class defines the interface that all SynthID watermark detectors must implement. Subclasses should override the detect() method to implement specific detection algorithms.

detect(g_values: numpy.ndarray, mask: numpy.ndarray, weights: numpy.ndarray = None) → numpy.ndarray

Computes the Weighted Mean score.

Args:

g_values: g-values of shape [batch_size, seq_len, watermarking_depth] mask: A binary array shape [batch_size, seq_len] indicating which g-values

should be used. g-values with mask value 0 are discarded

weights: array of non-negative floats, shape [watermarking_depth]. The

weights to be applied to the g-values. If not supplied, defaults to linearly decreasing weights from 10 to 1

Returns:

Weighted Mean scores, of shape [batch_size]. This is the mean of the unmasked g-values, re-weighted using weights.

detector_benchmark.watermark.synth_id.detector.get_detector(detector_name: str, logits_processor)