ODDN: Addressing Unpaired Data Challenges in Open-World Deepfake Detection on Online Social Networks

Renshuai Tao,Manyi Le,Chuangchuang Tan,Huan Liu,Haotong Qin,Yao Zhao

2024-10-24

Abstract:Despite significant advances in deepfake detection, handling varying image quality, especially due to different compressions on online social networks (OSNs), remains challenging. Current methods succeed by leveraging correlations between paired images, whether raw or compressed. However, in open-world scenarios, paired data is scarce, with compressed images readily available but corresponding raw versions difficult to obtain. This imbalance, where unpaired data vastly outnumbers paired data, often leads to reduced detection performance, as existing methods struggle without corresponding raw images. To overcome this issue, we propose a novel approach named the open-world deepfake detection network (ODDN), which comprises two core modules: open-world data aggregation (ODA) and compression-discard gradient correction (CGC). ODA effectively aggregates correlations between compressed and raw samples through both fine-grained and coarse-grained analyses for paired and unpaired data, respectively. CGC incorporates a compression-discard gradient correction to further enhance performance across diverse compression methods in OSN. This technique optimizes the training gradient to ensure the model remains insensitive to compression variations. Extensive experiments conducted on 17 popular deepfake datasets demonstrate the superiority of the ODDN over SOTA baselines.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to handle the challenges brought by unpaired data in open - world scenarios, especially on online social networks (OSN), in order to improve the performance of deep - fake detection. Specifically:

1. **Challenges of Unpaired Data**: Existing deep - fake detection methods usually rely on paired data (i.e., original images and compressed images). However, in the actual online social network environment, it is very difficult to obtain the original images, resulting in far more unpaired data than paired data. This data imbalance seriously affects the detection performance of existing methods.

2. **Compression Impact**: Online social network platforms usually perform various compression processes on uploaded images. These compression operations will reduce the image quality and mask the forgery traces, making it more difficult to identify forged images.

To solve these problems, the author proposes a new method named Open - World Deep - Fake Detection Network (ODDN), which contains two core modules:

- **Open - World Data Aggregation (ODA)**: It processes paired and unpaired data through fine - grained and coarse - grained analysis respectively, and effectively aggregates the correlation between compressed and original samples.

- **Compression - Dropout Gradient Correction (CGC)**: It introduces a compression - dropout gradient correction mechanism to optimize the training gradient and ensure the robustness of the model to different compression methods.

Through these two modules, ODDN aims to overcome the challenges brought by unpaired data and compression processing and improve the deep - fake detection performance in open - world scenarios. Experimental results show that ODDN outperforms the existing state - of - the - art methods on multiple datasets.