A Novel Blockchain-Based Deepfake Detection Method Using Federated and Deep Learning Models

Arash Heidari,Nima Jafari Navimipour,Hasan Dag,Samira Talebi,Mehmet Unal
DOI: https://doi.org/10.1007/s12559-024-10255-7
IF: 4.89
2024-01-26
Cognitive Computation
Abstract:Abstract In recent years, the proliferation of deep learning (DL) techniques has given rise to a significant challenge in the form of deepfake videos, posing a grave threat to the authenticity of media content. With the rapid advancement of DL technology, the creation of convincingly realistic deepfake videos has become increasingly prevalent, raising serious concerns about the potential misuse of such content. Deepfakes have the potential to undermine trust in visual media, with implications for fields as diverse as journalism, entertainment, and security. This study presents an innovative solution by harnessing blockchain-based federated learning (FL) to address this issue, focusing on preserving data source anonymity. The approach combines the strengths of SegCaps and convolutional neural network (CNN) methods for improved image feature extraction, followed by capsule network (CN) training to enhance generalization. A novel data normalization technique is introduced to tackle data heterogeneity stemming from diverse global data sources. Moreover, transfer learning (TL) and preprocessing methods are deployed to elevate DL performance. These efforts culminate in collaborative global model training zfacilitated by blockchain and FL while maintaining the utmost confidentiality of data sources. The effectiveness of our methodology is rigorously tested and validated through extensive experiments. These experiments reveal a substantial improvement in accuracy, with an impressive average increase of 6.6% compared to six benchmark models. Furthermore, our approach demonstrates a 5.1% enhancement in the area under the curve (AUC) metric, underscoring its ability to outperform existing detection methods. These results substantiate the effectiveness of our proposed solution in countering the proliferation of deepfake content. In conclusion, our innovative approach represents a promising avenue for advancing deepfake detection. By leveraging existing data resources and the power of FL and blockchain technology, we address a critical need for media authenticity and security. As the threat of deepfake videos continues to grow, our comprehensive solution provides an effective means to protect the integrity and trustworthiness of visual media, with far-reaching implications for both industry and society. This work stands as a significant step toward countering the deepfake menace and preserving the authenticity of visual content in a rapidly evolving digital landscape.
computer science, artificial intelligence,neurosciences
What problem does this paper attempt to address?
The problem that this paper attempts to solve is the detection and identification of deepfake videos. With the development of deep - learning technology, the production of deepfake videos has become increasingly easy. These videos can be almost indistinguishable from the real ones visually, thus posing a serious threat to the authenticity of the media. This threat not only affects news reporting and the entertainment industry, but may also be used for malicious purposes such as political deception and social manipulation. Therefore, effective technical means are required to detect and identify these deep - faked contents. To meet this challenge, this paper proposes a deep - fake detection method based on blockchain and federated learning (FL), combining the advantages of deep - learning models such as convolutional neural network (CNN) and capsule network (CN), aiming to improve the efficiency and generalization ability of image feature extraction. In addition, the paper also introduces a new data normalization technique to solve the problem of data heterogeneity from different global data sources. By using transfer learning (TL) and pre - processing methods, the performance of the deep - learning model is further improved. These measures work together to achieve efficient and accurate deep - fake video detection while protecting the privacy of data sources. Specifically, the main contributions of the paper include: - Proposing a framework named BFLDL, integrating edge computing and blockchain technology into the deep - fake detection method to ensure data protection and accuracy. - Proposing a data normalization method to correctly train the federated learning model with data from different sources. - Suggesting a lightweight deep - fake detection method, taking advantage of the excellent performance of the used deep - learning methods. - Using transfer learning to obtain the highest accuracy and area under the curve (AUC) while reducing the training time. In conclusion, the solution proposed in this paper not only improves the accuracy and efficiency of deep - fake video detection, but also ensures the security and privacy of data through blockchain and federated learning technologies.