Delving into the Local: Dynamic Inconsistency Learning for DeepFake Video Detection
Authors: Zhihao Gu, Yang Chen, Taiping Yao, Shouhong Ding, Jilin Li, Lizhuang Ma744-752
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our method outperforms the state of the art competitors on four popular benchmark dataset, i.e., Face Forensics++, Celeb-DF, DFDC and Wild Deepfake. Besides, extensive experiments and visualizations are also presented to further illustrate its effectiveness. |
| Researcher Affiliation | Collaboration | Zhihao Gu1,2*, Yang Chen2*, Taiping Yao2*, Shouhong Ding2 , Jilin Li2, Lizhuang Ma1,3,4 1School of Electronic and Electrical Engineering, Shanghai Jiao Tong University, 2You Tu Lab, Tencent 3Mo E Key Lab of Artificial Intelligence, Shanghai Jiao Tong University 4East China Normal University |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide any concrete access to source code or explicitly state its release. |
| Open Datasets | Yes | Datasets. We evaluate our method on four widely used benchmarks: Face Forensics++ (Rossler et al. 2019), DFDC (Dolhansky et al. 2019), Celeb-DF (Li et al. 2020b) and Wild Deepfake (Zi et al. 2020). |
| Dataset Splits | No | The paper mentions a 'validation set' ('We divide the learning rate by 10 when the performance on validation set saturates.') but does not provide specific details on the dataset splits (percentages, sample counts, or explicit splitting methodology). |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types, or memory amounts) used for running its experiments. |
| Software Dependencies | No | The paper mentions software components like 'dlib', 'MTCNN', 'Image Net pre-trained 2D Res Net-50', and 'Adam optimizer' but does not provide specific version numbers for any of them. |
| Experiment Setup | Yes | The image is resized to 224 224 during training. We adopt the Adam (Kingma and Ba 2014) as optimizer to optimize the binary cross-entropy loss. The batch size is 10 and the initial learning rate is 10 4. The total epoch is 30 for all datasets and 45 for cross-dataset generalization. We divide the learning rate by 10 when the performance on validation set saturates. Only horizontal flip is employed for augmentation. During inference, we sample U = 8 snippets with T = 4 frames and resize them into the same size as in training. |