RLDF: A Region-Level Deepfake Dataset and Benchmark for Partial-Face Manipulation Detection

Authors: Shukesh Reddy, Abhijit Das

Published: 2026-10-15 07:46:45+00:00

AI Summary

This paper introduces RLDF (Region-Level Deepfake), the first dataset for evaluating deepfake detection models on partial-face manipulations. It comprises over 2 million partial face-swap images, generated using eight face-swapping methods across seven facial regions, to address the limitation of existing datasets that focus solely on full-face manipulations. The study benchmarks 12 state-of-the-art detectors and proposes a mask-guided cross-attention model as a baseline, revealing that current detectors struggle with region-level deepfakes.

Abstract

Existing deepfake detection models are built upon manipu- lated images that are tampered with either exclusively fully swapped or fully synthesised faces, which leads us to an open question: Can off-the-shelf deepfake models detect manipulation effectively while tampering is confined to a local facial sub-region? To fill this gap, we introduce RLDF (Region-Level Deepfake), the first dataset for region-level deepfake evaluation. Employing the proposed region-aware partial swap synthesis model, we generate over 2M par- tial face-swap images by using eight face-swapping meth- ods. Altogether, we considered seven face regions, among which there are four discrete regions (eye, nose, mouth and eyebrow) and four composite facial regions (eye and nose; eye, nose and mouth; eye, nose, mouth and eye- brow; and full face). In addition, we define three evalu- ation protocols and benchmark 12 state-of-the-art detectors across the five popular standard cross-dataset evaluations, DF40, and the proposed RLDF, providing a comprehen- sive audit of where and why existing detectors underper- formed. Our evaluation reveals that existing detectors fail to generalize consistently across partial facial regions, re- sulting in an overall drop of 10.5% AUC and 10% accu- racy under region-level evaluation. Further, we propose a mask-guided cross-attention model as a baseline approach that can effectively detect region-level manipulation. The proposed dataset, code, and checkpoints are available at https://github.com/RLDF-Dataset/RLDF.

Key findings

Existing deepfake detectors fail to generalize consistently across partial facial regions, showing an average drop of 10.5% AUC and 10% accuracy under region-level evaluation. The proposed mask-guided cross-attention model effectively detects region-level manipulation by focusing on patch-level forgery grounding, demonstrating a generalization advantage. The three evaluation protocols (classic cross-dataset, DF40, and RLDF region-level) are orthogonal, indicating that strong performance on one does not predict sensitivity on others.

Approach

The authors generate a new dataset, RLDF, by applying existing face-swapping methods to specific facial sub-regions using a region-aware partial swap synthesis model. They then benchmark 12 state-of-the-art deepfake detectors on this dataset and propose a mask-guided cross-attention model that leverages CLIP ViT-L/14 to predict both a binary label and a spatial forgery map, guiding attention to suspicious patches.

Datasets

RLDF (proposed), FaceForensics++, CelebDF-v1/v2, DFDC, DFDCP, UADFV, DF40

Model(s)

Xception, F3Net, RECCE, SBI, LGrad, UCF, FatFormer, FreqNet, ProDet, DFD-FCG, Effort, ForAda, CLIP ViT-L/14 (for proposed mask-guided cross-attention model)

Author countries

India

← Previous