Existing deepfake detection models are typically trained on images where the entire face has been swapped or synthesised. That leaves a gap: many real-world manipulations alter only a region of the face, such as the mouth or eyes, and it is unclear whether current detectors can catch such edits.

The paper introduces RLDF, a region-level deepfake dataset and benchmark designed specifically for partial-face manipulation detection. By focusing on tampered regions rather than whole faces, RLDF aims to test the generalisation of off-the-shelf deepfake models to a more realistic and challenging setting.

The abstract frames the central question directly: can off-the-shelf deepfake models detect these partial manipulations? The benchmark is intended to provide a standard way to answer that question, highlighting a potential blind spot in existing detection systems.