Dynamic Inconsistency-aware DeepFake Video Detection

Authors: Ziheng Hu, Hongtao Xie, YuXin Wang, Jiahong Li, Zhongyuan Wang, Yongdong Zhang

IJCAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In experiments, we quantitatively verify the effectiveness of utilizing inconsistency information between adjacent frames. Benefiting from exploring the global and local inconsistencies, our model outperforms existing methods on Face Forensics++ (FF++) [Rossler et al., 2019], DFDC-preview [Dolhansky et al., 2019] and shows a good rubustness to degradation of video quality and unseen manipulation techniques.
Researcher Affiliation Collaboration 1University of Science and Technology of China 2Kuaishou Technology {hzh519, wangyx58}@mail.ustc.edu.cn, {htxie, zhyd73}@ustc.edu.cn, {lijiahong, wangzhongyuan}@kuaishou.com
Pseudocode No The paper describes the proposed method and its components using text, diagrams, and mathematical equations, but does not include any explicitly labeled
Open Source Code No The paper does not provide an explicit statement or link indicating that the source code for the described methodology is publicly available.
Open Datasets Yes To verify the effectiveness and generalization of the proposed DIANet, we conduct experiments on multiple datasets: Face Forensics++ (FF++) [Rossler et al., 2019], Celeb-Deep Fake v2 (Celeb-DFv2) [Li et al., 2020b] and Deep Fake Detection Challenge preview (DFDC-preview) [Dolhansky et al., 2019].
Dataset Splits No The paper states
Hardware Specification No The paper does not specify any particular GPU models, CPU models, or other detailed hardware specifications used for running the experiments.
Software Dependencies No The paper mentions using
Experiment Setup Yes The input images are resized to 3 224 224. The dimension of the feature representation V1 and V2 is 256 29 29. The output dimension of Ro IPooling (i.e. the size of W1 and L1) is 256 14 14. The GCB and LAB are trained with SGD. The SGD optimizer is used with an initial learning rate of 1 10 3 with momentum of 0.9 and weight decay of 1 10 4. During training, the batchsize is set to 32.