Organizations: LARIVIA laboratory Dep. of Computer science and Eng. University of Quebec in Outaouais Gatineau, QC, Canada · LARIVIA laboratory and Mixed Research Unit INRS-UQO University of Quebec in Outaouais Gatineau, QC, Canada
Deepfake detection in videos remains challenging, as manipulated content may appear visually consistent at the frame level while exhibiting subtle temporal inconsistencies. This paper introduces an interpretable deepfake detection framework that models spatially and temporally coherent facial features in video sequences. Unlike end-to-end deep models relying on implicit representations, the proposed approach explicitly encodes physically grounded forensic cues, enabling transparent analysis and improved multi-dataset generalization. The pipeline transforms videos into identity-consistent facial trajectories, segments them into fixed-length temporal windows, and represents each frame using 68 structured descriptors spanning four complementary domains: photometric, textural, geometric, and compression-based features. These descriptors provide a compact multi-domain representation of manipulation artifacts and are processed by a Long Short-Term Memory (LSTM) network to capture temporal dependencies and subtle irregularities. Evaluation on four benchmark datasets, FaceForensics++, Celeb-DF v2, a curated subset of the DeepFake Detection Challenge (DFDC), and DeeperForensics, yields strong and consistent F1-scores of 98.0%, 91.0%, 97.6%, and 96.2%, respectively. The approach also demonstrated a good cross-dataset generalization, providing a robust and interpretable solution for video deepfake detection.
Figures & tables
Fig. 1: Facial analyzed area: (Left) Original input face image, (Middle) Segmented facial region (green overlay) where color and texture features are extracted. Automatically detected landmarks (red dots) used to compute geometric features. The midline divides the face into left (L) and right (R) halves, enabling quantitative assessment of facial symmetry. (Right) is the analyzed regions covering hair-skin contact and jaws.
Photometric
Textural
Geometric
Compression
Idx
Variable
Idx
Variable
Idx
Variable
Idx
Variable
1
mean_R
19
GLCM_contrast_L
43
inter_eye_dist
57
ELA_mean_face
2
mean_G
20
GLCM_contrast_R
44
nose_to_mouth_dist
58
ELA_std_face
3
mean_B
21
GLCM_corr_L
45
mouth_width
59
ELA_mean_mouth
4
std_R
22
GLCM_corr_R
46
mouth_opening
60
ELA_std_mouth
5
std_G
23
GLCM_energy_L
47
face_ratio
61
jpeg_res_face
TABLE I: Complete Feature Set Organized by Forensic Families
Dataset
Acc.
Prec.
Rec.
F1
AUC
FaceForensics++
98.0%
97.9%
98.1%
98.0%
0.995
Celeb-DF v2
91.0%
90.8%
91.2%
91.0%
0.948
DFDC subset
97.6%
97.4%
97.8%
97.6%
0.991
DeeperForensics
96.2%
96.0%
96.4%
96.2%
0.984
Average
95.7%
95.5%
95.9%
95.7%
0.980
TABLE II: Dataset-wise performance of the proposed method. The last row reports the unweighted arithmetic average across datasets, not the global test-set performance.
Predicted Real
Predicted Fake
Real
982
18
Fake
24
976
TABLE III: Illustrative Confusion Matrix on a Balanced Test Subset
Method
Year
Method type / Protocol
Reported quantitative results
Avg. AUC
Amin et al. [ 19 ]
2024
Temporal facial pattern and coherence analysis; evaluated on FF++, DFD, DFDC, DeeperForensics
TABLE IV: Quantitative comparison with recent deepfake detection methods. The table reports only the metrics explicitly provided by the corresponding papers. Since protocols, datasets, and evaluation settings differ across works, this table is intended as a quantitative reference comparison rather than a strict leaderboard.
Dataset
Acc.
Prec.
Rec.
F1
AUC
ForgeryNet
93.8%
93.5%
94.1%
93.8%
0.967
WildDeepfake
89.4%
88.8%
90.2%
89.5%
0.941
TABLE V: Cross-dataset generalization results. ForgeryNet is used as validation and WildDeepfake as test. No retraining is performed.
Dataset
Ground Truth
Pred. Real
Pred. Fake
ForgeryNet
Real
935
65
Fake
59
941
WildDeepfake
Real
887
113
Fake
98
902
TABLE VI: Illustrative confusion matrices for cross-dataset evaluation.
Configuration
Acc.
Prec.
Rec.
F1
AUC
All 4 families
93.8%
93.5%
94.1%
93.8%
0.967
Without photometric
90.9%
90.5%
91.3%
90.9%
0.941
Without textural
91.6%
91.2%
92.0%
91.6%
0.948
Without geometric
92.0%
91.7%
92.4%
92.0%
0.953
Without compression
90.6%
90.1%
91.1%
90.6%
0.938
TABLE VII: Ablation on feature families — ForgeryNet (Validation).
Configuration
Acc.
Prec.
Rec.
F1
AUC
All 4 families
89.4%
88.8%
90.2%
89.5%
0.941
Without photometric
85.8%
85.0%
86.9%
85.9%
0.902
Without textural
86.7%
86.0%
87.6%
86.8%
0.913
Without geometric
87.2%
86.6%
88.0%
87.3%
0.919
Without compression
85.4%
84.6%
86.6%
85.6%
0.899
TABLE VIII: Ablation on feature families — WildDeepfake (Test).
Fig. 2: Distribution of selected photometrc, geometric, textural, and compression-related forensic features for real (orange) and fake (blue) samples extracted from the FaceForensic++ dataset.
Shenzhen Technology University, Guangdong, China · Singapore Institute of Technology, Singapore, Singapore · The University of Hong Kong, Hong Kong, China +1
GS. of AI, Chung-Ang University, Republic of Korea · Dept. of Advanced Imaging, GSAIM, Chung-Ang University, Republic of Korea · Dept. of Metaverse Convergence, Chung-Ang University, Republic of Korea