Explainable Deepfake Detection
Overview
Explainable deepfake detection studies how to identify manipulated images, videos, and audio while making the evidence behind a prediction understandable and checkable. Approaches include highlighting suspicious regions or time segments, using explicit forensic and phonetic features, and training vision-language models to connect explanations to localized evidence. Evaluation uses datasets and benchmarks for detection, localization, explanation quality, and cross-dataset testing across generators, recording conditions, and media types. A central challenge is ensuring that convincing explanations reflect cues that actually drive the decision, while remaining reliable on unfamiliar manipulations and after compression or other changes.
Latest papers
63 papers8 papers in the last four weeks, down 20% on the four weeks before. 0.1% of all new papers.