Applying Encoding-Decoding Direction Pairs to an Xception deepfake detector reveals 16 interpretable concepts (e.g., fake-mouth, real-eyes) that drive real/fake predictions, with concept-level interventions achieving 99.8% correction of misclassified samples.
Self-supervised learning of adversarial example: Towards good generalizations for deepfake detection
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Why Fake ? Unveiling the Semantic Vocabulary of Deepfake Detectors
Applying Encoding-Decoding Direction Pairs to an Xception deepfake detector reveals 16 interpretable concepts (e.g., fake-mouth, real-eyes) that drive real/fake predictions, with concept-level interventions achieving 99.8% correction of misclassified samples.