Extending the training-free embedding-dynamics paradigm beyond audio

Extend the training-free embedding-dynamics paradigm used for TRACE to other modalities by constructing and evaluating analogous detectors for deepfake face detection with vision transformers, machine-generated text detection with language models, and cross-modal consistency verification in multimodal foundation models.

Background

TRACE demonstrates that first-order dynamics of frozen speech foundation model embeddings can detect partial audio deepfakes without training or labeled data, suggesting a modality-agnostic forensic signal in foundation model geometry.

The authors explicitly state that extending this paradigm beyond audio—to images/videos, text, and multimodal consistency—remains an open direction for future research.