Collaborative Decentralized Learning for Detecting Deepfake Videos in Entertainment

In this work, we propose decentralized collaborative learning adopted for detecting deepfake videos in entertainment. Most state-of-the-art works in this field utilize a centralized learning mechanism of Artificial Intelligence (AI), which may pose privacy concerns for participants using actual image datasets, as data needs to be forwarded to a central server. To mitigate this, Federated Learning (FL) has been developed. In this work, the aggregation mechanism is performed in a decentralized manner for enhanced security, as the centralized FL approach is prone to single points of failure. The proposed decentralized collaborative learning is trained by each participant using real and deepfake video datasets. These videos are sampled and trained individually before being forwarded to other participants in the proposed collaborative system. In this work, continual learning mechanisms are employed to reduce the communication overhead required during the aggregation process. A total of 953 videos, each with an average duration of 10 seconds, were utilized for training and evaluation. The results demonstrate that the proposed decentralized collaborative learning approach can achieve satisfactory performance, with an accuracy of 97.30%, while also preserving participant privacy and minimizing communication costs.