FInfer: Frame Inference-Based Deepfake Detection for High-Visual-Quality Videos

Authors: Juan Hu, Xin Liao, Jinwen Liang, Wenbo Zhou, Zheng Qin951-959

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate the performance of our method is promising in terms of in-dataset detection performance, detection efficiency, and cross-dataset detection performance in high-visualquality Deepfake videos.
Researcher Affiliation Academia 1The College of Computer Science and Electronic Engineering, Hunan University, Changsha 410082, China. 2CAS Key Laboratory of Electromagnetic Space Information, University of Science and Technology of China.
Pseudocode Yes Algorithm 1: The algorithm process of the proposed frame inference-based detection framework.
Open Source Code No The paper does not provide any specific links to a code repository or explicitly state that the source code for the described methodology is publicly available.
Open Datasets Yes The Face Forensics++ (FF++) (Rossler et al. 2019) dataset, the Celeb-DF dataset (Li et al. 2020), the Wild Deepfake dataset (Zi et al. 2020), and the DFDCpreview dataset(Dolhansky et al. 2019) are utilized to show the performance of FInfer.
Dataset Splits No The paper analyzes the impact of parameters 's' and 't' on detection accuracy on the Celeb-DF dataset, implying a form of validation. However, it does not explicitly provide information on train/validation/test splits as percentages, absolute counts, or references to predefined standard splits for overall model training and evaluation.
Hardware Specification Yes All experiments are conducted in Keras on NVIDIA Titan Xp.
Software Dependencies No The paper mentions software like FFmpeg, dlib, Keras, and Adam optimizer but does not specify version numbers for these dependencies, which would be necessary for reproducible setup.
Experiment Setup Yes The batch size is set as 8. In the training phase, we set the learning rate as 0.001, which will be divided by 5 when the accuracy plateaus. The Adam optimizer (Kingma and Ba 2014) is utilized to optimize the model. We set the default threshold, whose value is 0.5, to calculate binary accuracy.