Distribution Shift
Definition
In machine learning, distribution shift occurs when the statistical characteristics of data the model encounters in deployment differ from those of the data it was trained on. For deepfake detectors, training on one generation of generative models and deploying against a newer generation is a common and serious form of distribution shift.
- Field
- Machine learning / media forensics
- Definition
- Deployment data statistics differ from training data statistics
- Common cause in deepfake work
- Newer generative model versions than the training set
- Consequence
- Degraded detector accuracy in the field
Common questions
Why can a deepfake detector that scored well in testing fail in real casework?+
If the detector was trained and validated on outputs from older generative models, it can miss or misclassify content from newer generators whose artefacts differ, so a high reported accuracy does not guarantee performance on current media.
How do labs guard against distribution shift in deployed detection tools?+
Ongoing revalidation against current-generation synthetic content, rather than a one-time accuracy figure, is the main safeguard, along with reporting a detector's confidence as bounded by its training distribution rather than as a categorical result.
Related terms
- C2PA
- Coalition for Content Provenance and Authenticity. A cross-industry group that has published an open technical specification for embedding cryptographically signed provenance manifests...
- Daubert Standard
- The US federal evidentiary standard (Daubert v. Merrell Dow Pharmaceuticals, 1993) requiring that expert testimony be based on scientifically valid methods with...
- ISO/IEC 17025
- The international standard for testing and calibration laboratories, published jointly by the International Organization for Standardization and the International Electrotechnical Commission. It...
- NIST CFTT
- The National Institute of Standards and Technology Computer Forensics Tool Testing programme. It publishes independent test reports for digital forensic tools, including...
- SWGDE
- Scientific Working Group for Digital Evidence. A US multi-agency body that publishes consensus best-practice documents for digital forensic disciplines, including image authentication,...