Leakage-Aware Deepfake Video Detection and Explainable Verification: Dataset Auditing, Split Design and Cross-Dataset Evaluation

Deepfake detectors often score well on benchmarks but degrade on videos from other datasets, manipulation methods and compression levels. This work studies deepfake video detection with emphasis on cross-dataset generalization, dataset integrity and reliable evaluation, using EfficientNet-B4 and Xception classifiers on FaceForensics++ (C23), Celeb-DF and additional collected videos. An audit of 17,407 video records (14,888 fake; 2,519 real) found zero byte-identical duplicate groups, but an earlier filename-derived split showed about 90.7% filename-token overlap between test and train/validation rows. We construct a revised split (V11) and a sensitivity-analysis candidate (V12) that groups seven candidate visual links within single partitions, with zero cross-split pairs among 76,263 existing visual-match rows. The V11 model reached 86.69% accuracy and 74.76% AUROC on the external Celeb-DF set (6,528 videos), but balanced accuracy was only 51.65% and real-class specificity 3.49%. V12 results reported are validation-only (best AUROC 0.981005 at epoch 2); V12 test and Celeb-DF evaluation are pending. The findings show that accuracy and fake-class recall alone are insufficient evidence of reliable detection.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-20
DOI
https://doi.org/10.5281/zenodo.24386320
Primary Topic
Digital Media Forensic Detection
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Leakage-Aware Deepfake Video Detection and Explainable Verification: Dataset Auditing, Split Design and Cross-Dataset Evaluation

Soumyadeep Ghosh, Ngangbam Indrason, Shaurya Yadav, Bidyashore LOUREMBAM BIDYASHORE et al.
3 links
Zenodo (CERN European Organization for Nuclear Research)
Digital Media Forensic Detection
preprint

Leakage-Aware Deepfake Video Detection and Explainable Verification: Dataset Auditing, Split Design and Cross-Dataset Evaluation

Soumyadeep Ghosh, Ngangbam Indrason, Shaurya Yadav, Bidyashore LOUREMBAM BIDYASHORE, Ketan AGARWAL Agrawal
preprint en

Abstract

Deepfake detectors often score well on benchmarks but degrade on videos from other datasets, manipulation methods and compression levels. This work studies deepfake video detection with emphasis on cross-dataset generalization, dataset integrity and reliable evaluation, using EfficientNet-B4 and Xception classifiers on FaceForensics++ (C23), Celeb-DF and additional collected videos. An audit of 17,407 video records (14,888 fake; 2,519 real) found zero byte-identical duplicate groups, but an earlier filename-derived split showed about 90.7% filename-token overlap between test and train/validation rows. We construct a revised split (V11) and a sensitivity-analysis candidate (V12) that groups seven candidate visual links within single partitions, with zero cross-split pairs among 76,263 existing visual-match rows. The V11 model reached 86.69% accuracy and 74.76% AUROC on the external Celeb-DF set (6,528 videos), but balanced accuracy was only 51.65% and real-class specificity 3.49%. V12 results reported are validation-only (best AUROC 0.981005 at epoch 2); V12 test and Celeb-DF evaluation are pending. The findings show that accuracy and fake-class recall alone are insufficient evidence of reliable detection.

Zenodo (CERN European Organization for Nuclear Research)
VIT-AP University (IN), Visvesvaraya Technological University (IN)
Digital Media Forensic Detection
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon