[공지사항]을 빙자한 안부와 근황 
Show more

Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Uncovering Critical Features for Deepfake Detection through the Lottery Ticket Hypothesis

Created by
  • Haebom

Author

Lisan Al Amin, Md. Ismail Hossain, Thanh Thi Nguyen, Tasnim Jahan, Mahbubul Islam, Faisal Quader

Outline

In this paper, we propose an efficient deepfake detection method using the Lottery Ticket Hypothesis (LTH) to address the threats to information integrity and social trust caused by the rapid increase in deep synthetic media. We conduct experiments by applying architectures such as MesoNet, CNN-5, and ResNet-18 to the OpenForensic and FaceForensics++ datasets, and show that it is possible to reduce the model size without performance degradation through neural network pruning. In particular, in the case of MesoNet, we achieve an accuracy of 56.2% on the OpenForensic dataset (about 90% compared to the baseline accuracy of 62.6%) even at a pruning ratio of 80%, and we obtain results using only 3,000 parameters. In addition, we show that the proposed LTH-based iterative size pruning method is superior to the batch pruning method, and we confirm that the pruned network focuses on the face region that is important for deepfake detection through Grad-CAM. We also show the transferability of winning tickets between datasets, suggesting the possibility of an efficient and deployable deepfake detection system.

Takeaways, Limitations

•
Takeaways:
◦
We experimentally demonstrate that LTH can effectively reduce the size of deepfake detection networks.
◦
Presenting the possibility of building an efficient deepfake detection system even in resource-constrained environments.
◦
We verify that the pruned network still maintains detection performance by focusing on important facial features.
◦
The transferability of winning tickets suggests applicability to various datasets.
◦
It is demonstrated that the iterative size pruning method is superior to the batch pruning method.
•
Limitations:
◦
Further research is needed on the generalization performance of the proposed method.
◦
Additional validation is needed on various deepfake generation techniques and datasets.
◦
Further research is needed on performance evaluation and deployment in real environments.
👍
Made with Slashpage