Original Research (Published On: 16-Nov-2026 )
DOI : https://doi.org/10.54364/AAIML.2026.66358Su-Cheng Haw, Jayapradha J, Senthil Kumar T, Malathi S and Nilanjana Bhattacharya
Adv. Artif. Intell. Mach. Learn., - (-):-
1. Su-Cheng Haw: Faculty of Computing and Informatics, Multimedia University, Persiaran Multimedia 63100, Cyberjaya, Selangor
2. Jayapradha J: Department of Computing Technologies, School of Computing, SRM Institute of Science and Technology, Kattankulathur, Tamil Nadu 603203, India
3. Senthil Kumar T: Department of Computing Technologies, School of Computing, SRM Institute of Science and Technology, Kattankulathur, Tamil Nadu 603203, India
4. Malathi S: Department of Electrical and Electronics Engineering, SASTRA University, Thanjavur, India
5. Nilanjana Bhattacharya: Software Development Engineer-Level 1, JP Morgan Chase and Co, Hyderabed, India
DOI: 10.54364/AAIML.2026.66358
Article History: Received on: 20-Jul-26, Accepted on: 09-Nov-26, Published on: 16-Nov-26
Corresponding Author: Su-Cheng Haw
Email: sucheng@mmu.edu.my
Citation: Jayapradha J, et al. Enhancement of Deepfake Identification Using Spatial, Temporal, Frequency and Audio-Visual Fusion Techniques. Advances in Artificial Intelligence and Machine Learning. 2026. [Ahead of Print]. https://dx.doi.org/10.54364/AAIML.2026.66358
Abstract
Rapid advancements in
deepfake generation technology have met challenges in media. Dependability
truthfulness authentication. Therefore, this paper presents a multi-modal
deepfake detection anti-cheat system: spatial, temporal, frequency, and
audiovisual synchrony features were incorporated. Visualization of ResNet18
pre-trained on imagenet data pounds down pixel-level features in the spatial
channel, while a Long Short-Term Memory (LSTM) recurrent neural network in the
temporal channel captures sequential behavior. Frequency artifacts are
extracted using Fast Fourier Transform (FFT) and audiovisual synchrony compared
by Dynamic Time Warping (DTW) between Mel Frequency Cepstral Coefficients
(MFCC, audio) and lip movement locations (visual). In the last step, we unify
spatial, temporal, frequency, and audio–visual synchrony features. The four
feature streams are concatenated into a 1,153-dimensional vector for
classification. The manuscript reports accuracies of 78.4%, 82.1%, and 79.6%
for the spatial, temporal, and frequency models, and 91.3% for the fused model,
with precision 0.901, recall 0.918, and F1-score 0.910 for fusion.Experimentally,
a model is proposed for real-world testing that applies YouTube data as input
to compare real and manipulated media without time limits. With multi-modal
deepfake detection framework, cross spatial, temporal, frequency and
audio-visual synchrony detection is of a higher-class stability and reduces the
misjudgments rate.
Statistics
Article Views: 1
PDF Downloads: 0
