Limited generalization is the inability of a detection method or model to perform reliably on unfamiliar data or manipulation types. A detector may work well on known manipulations but fail when novel forms are absent from its training data. section 4.4
In synthetic-media detection, limited generalization creates several problems:
- Newer generative models can produce diverse, customizable manipulations that detectors have not encountered. section 4.4
- Detection becomes adversarial: each successful technique can encourage generative models to adapt and evade it. section 4.4
- Visual or auditory analysis alone cannot reliably identify every form of synthetic media at scale. section 4.4
Synthetic-media detection
A detector trained on familiar deepfake patterns may perform well in testing but struggle with a new manipulation absent from its training data. section 4.4
