Generalisation Gap
Definition
The drop in detection accuracy when a classifier trained on one generation method is applied to a different method. Caused by learning artifacts specific to a particular GAN or pipeline rather than general forgery indicators. The central open problem in deepfake detection.
- Domain
- Deepfake and GAN-image detection
- Cause
- Learned method-specific artifacts, not general forgery cues
- Effect
- Accuracy drop on unseen generation methods
- Status
- Unsolved, central open problem in the field
Common questions
Why does a detector trained on one deepfake method fail on another?+
Detectors often learn to recognize artifacts specific to a particular generator or pipeline rather than features common to all manipulated content, so an unseen generation method produces different artifacts the classifier never learned.
How do researchers try to reduce the generalisation gap?+
Approaches include training on outputs from diverse generators, using frequency-domain features less tied to a specific architecture, and foundation-model detectors trained on broad data distributions.
Does a high accuracy score on one benchmark guarantee real-world performance?+
No, a detector can score well on same-method test data while performing near chance on deepfakes from an unseen generator, which is why cross-dataset evaluation is now standard practice.
Related terms
- CLIP-Based Detection
- Detection approach using OpenAI CLIP or similar vision-language foundation models as a feature extractor. The broad pre-training enables generalisation to generation methods...
- CNN Residual Detector
- A convolutional neural network trained on the high-frequency residual image, the difference between the original and a de-noised version, to classify whether...
- Detection Generalisation
- The capacity of a trained detector to correctly identify deepfakes produced by generators not seen during training. Low generalisation is the central...
- FaceForensics++
- A video dataset released by Rossler et al. (2019) containing 1000+ YouTube videos manipulated by four methods: DeepFakes, Face2Face, FaceSwap, and NeuralTextures....
- Frequency-Domain Analysis
- Detection approach that transforms image patches into the frequency domain (DCT or FFT) to expose periodic artifacts introduced by upsampling layers in...
- Frequency-Domain Artefact
- A periodic or statistical anomaly in the Fourier spectrum of an image or audio signal introduced by the generation pipeline's upsampling, filter,...
- Noiseprint
- A CNN-based camera-model fingerprint extractor by Cozzolino and Verdoliva. Applied to deepfakes, it reveals inconsistency between the camera fingerprint in the genuine...
- Physiological Signal
- A biological process visible in video, such as eye blinking, rPPG (remote photoplethysmography), and head micro-motion from the cardiac cycle, that deepfake...
- Remote Photoplethysmography (rPPG)
- A technique that detects the pulse-driven skin-colour variation in a face video without contact sensors. In real video, this signal is present...
- rPPG
- Remote photoplethysmography. A technique for measuring heart rate from subtle periodic colour changes in facial skin caused by blood-volume pulses. Authentic video...
- XceptionNet
- A depthwise-separable convolutional architecture proposed by Rossler et al. as the baseline binary classifier in FaceForensics++. Trained to distinguish real from manipulated...