Proposal / Submission Type
Peer Reviewed Paper
Abstract
Deepfake has brought huge threats to society such that everyone can become a potential victim. Current Deepfake detection approaches have unsatisfactory performance in either accuracy or efficiency. Meanwhile, most models are only evaluated on different benchmark test datasets with different accuracies, which could not imitate the real-life Deepfake unknown population. As Deepfake cases have already been raised and brought challenges at the court, it is disappointed that no existing work has studied the model reliability and attempted to make the detection model act as the evidence at the court. We propose a lightweight Deepfake detection deep learning approach using the convolutional neural network backbone and the efficient convolutional attention mechanism, outperforming the state-of-the-art baseline models on each benchmark test dataset. Furthermore, a real-life Deepfake content is usually unknown about the corresponding source dataset or manipulation technique. We conduct a model reliability study using statistical random sampling from the available benchmark datasets to imitate the real-life Deepfake cases. A sufficient number of trials for model evaluation with random sampling derives the 95% and 90% confidence intervals, informing the reliable accuracy information of the proposed model. As a result, the reliably quantified detection model derives satisfactory accuracy and error rate to be applicable at the court for civil cases and provides an informative scheme to analyze future satisfactory approaches for criminal cases at the court.
Scholarly Commons Citation
Wang, Tianyi and Chow, Kam Pui, "A Lightweight Reliably Quantified Deepfake Detection Approach" (2022). Annual ADFSL Conference on Digital Forensics, Security and Law. 8.
https://commons.erau.edu/adfsl/2022/presentations/8
Included in
Aviation Safety and Security Commons, Computer Law Commons, Defense and Security Studies Commons, Forensic Science and Technology Commons, Information Security Commons, National Security Law Commons, OS and Networks Commons, Other Computer Sciences Commons, Social Control, Law, Crime, and Deviance Commons
A Lightweight Reliably Quantified Deepfake Detection Approach
Deepfake has brought huge threats to society such that everyone can become a potential victim. Current Deepfake detection approaches have unsatisfactory performance in either accuracy or efficiency. Meanwhile, most models are only evaluated on different benchmark test datasets with different accuracies, which could not imitate the real-life Deepfake unknown population. As Deepfake cases have already been raised and brought challenges at the court, it is disappointed that no existing work has studied the model reliability and attempted to make the detection model act as the evidence at the court. We propose a lightweight Deepfake detection deep learning approach using the convolutional neural network backbone and the efficient convolutional attention mechanism, outperforming the state-of-the-art baseline models on each benchmark test dataset. Furthermore, a real-life Deepfake content is usually unknown about the corresponding source dataset or manipulation technique. We conduct a model reliability study using statistical random sampling from the available benchmark datasets to imitate the real-life Deepfake cases. A sufficient number of trials for model evaluation with random sampling derives the 95% and 90% confidence intervals, informing the reliable accuracy information of the proposed model. As a result, the reliably quantified detection model derives satisfactory accuracy and error rate to be applicable at the court for civil cases and provides an informative scheme to analyze future satisfactory approaches for criminal cases at the court.
