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

TL;DR

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), which effectively aggregates correlations between compressed and raw samples through both fine-grained and coarse-grained analyses for paired and unpaired data.

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.

Figures (4)

• Figure 1: Comparison of training data and models between traditional and open-world scenarios. Open-world scenarios require models to effectively manage the unpaired data sources, addressing the inherent variability and complexity.
• Figure 2: Overview of the proposed Open-world Deepfake Detection Network (ODDN). The ODDN contains two core modules: Open-world Data Aggregation (ODA) and compression-discard Gradient Correction (CGC). The ODA module focuses on aligning unpaired data in open-world OSN by leveraging the feature-center points. The CGC module ensures the proper optimization direction through integrated gradient correction during the process of removing compression-sensitive information.
• Figure 3: Performance comparison across 17 well-known datasets in quality-agnostic experiments (simulating the open-world OSN scenario) is illustrated for both the 2-class (left figure) and 4-class (right figure) training settings. The ODDN we proposed outperforms these baselines overall.
• Figure 4: The feature visualization of baseline and ODDN.