EmbeddingCopyrightDetector<T>

Models & Types · AiDotNet.Safety.Text

Detects potential copyright violations using embedding-based semantic similarity to known copyrighted works.

Inheritance: ModelBase<T, Vector<T>, Vector<T>> → SafetyModuleBase<T> → TextSafetyModuleBase<T> → EmbeddingCopyrightDetector<T>

For Beginners

If someone rewrites a copyrighted book using different words but the same ideas and structure, n-gram matching won’t catch it. This module converts text into mathematical representations that capture meaning, so it can detect “same content, different words” situations.

How It Works

While NgramCopyrightDetector catches verbatim copying, this detector catches paraphrased or semantically similar reproductions of copyrighted content. Each reference work is split into passages, embedded into a fixed-dimensional vector space, and compared against the model output using cosine similarity. High similarity to specific passages indicates potential memorization even when exact wording differs.

References:

  • DE-COP: Detecting copyrighted content via paraphrased permutations (2024, arxiv:2402.09910)
  • CopyBench: Measuring literal and non-literal copyright memorization (2024, arxiv:2407.07087)
  • BookMIA: Practical membership inference for book-level memorization (2024, arxiv:2401.15588)

Constructors

Constructor Summary
EmbeddingCopyrightDetector(string[] copyrightedTexts, string[] sourceNames, double threshold, int embeddingDim, int passageLength) Initializes a new embedding-based copyright detector.

Properties

Property Summary
string ModuleName { get; }

Methods

Method Summary
IReadOnlyList<SafetyFinding> EvaluateText(string text)