Deepfake Video Detection via Facial Action Dependencies Estimation

Authors: Lingfeng Tan, Yunhong Wang, Junfu Wang, Liang Yang, Xunxun Chen, Yuanfang Guo

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate the superiority of our method against the state-of-the-art methods. ... Experiments Setups Datasets The training of our models is carried out on the Face Forensics++ (FF++) (R ossler et al. 2018) dataset. ... Ablation Study In this subsection, we conduct a series of ablation studies on different components of FADE.
Researcher Affiliation Collaboration 1 School of Computer Science and Engineering, Beihang University, China 2 School of Artificial Intelligence, Hebei University of Technology, China 3 CNCERT/CC, Beijing, China 4 Zhongguancun Laboratory, Beijing, China
Pseudocode No The paper describes the method using equations and architectural diagrams, but does not provide explicit pseudocode or algorithm blocks.
Open Source Code No The paper does not mention providing open-source code or a link to a code repository for the described methodology.
Open Datasets Yes The training of our models is carried out on the Face Forensics++ (FF++) (R ossler et al. 2018) dataset. ... Celeb-DF (CDF) (Li et al. 2020c) and Deep Fake Detection (DFD) (Nick Dufour and Andrew Gully 2019).
Dataset Splits Yes For each video, four clips with the same number of frames are evenly sampled for training, validation and testing.
Hardware Specification Yes These models are trained on four RTX 3080s for approximately 10000 iterations and the best model is selected according to the evaluations on the validation set.
Software Dependencies No The paper mentions optimizers (Adam W) and a backbone network (Xception) but does not specify version numbers for any software dependencies.
Experiment Setup Yes 8 continuous frames with an interval of 8 are sampled from the videos as the input. ... In the training process, the batch size is set as 4 (clips). The initial learning rate of MDGM is 2e-4, while the learning rate of the backbone is one fifth of it. The learning rates will be divided by 2 when the AUC score plateaus for 3 epochs. The models are optimized via Adam W (Loshchilov and Hutter 2017) optimizer with β1 = 0.9, β2 = 0.999, and ϵ = 0.001.