DPL: Cross-Quality DeepFake Detection via Dual Progressive Learning

Dongliang Zhang, Yunfei Li, Jiaran Zhou, Yuezun Li

Introduction

As DeepFake generation techniques become increasingly sophisticated, the need for robust detection systems has never been more urgent. However, a significant gap exists between laboratory performance and real-world application. Most existing detectors are trained on high-quality, uncompressed datasets, yet real-world videos—especially those on social media platforms like TikTok, YouTube, or X—undergo heavy lossy compression. This compression "smears" the subtle digital artifacts that detectors rely on, making forgery patterns nearly indistinguishable from natural image noise.

To address this, the paper introduces Dual Progressive Learning (DPL), a framework that treats DeepFake detection as a dynamic process of "excavation." Rather than applying a static amount of computation to every video, DPL adaptively "drills" deeper into lower-quality or more complex samples to uncover concealed forgery traces.

The Challenge of Cross-Quality Detection

Detecting forgeries in compressed videos is difficult because the compression process effectively acts as a low-pass filter, removing high-frequency details where many forgery signatures reside. Previous attempts to solve this "cross-quality" problem often involved training multiple models for different quality levels or using complex ensembles. These approaches are computationally expensive and often fail when the quality gap between training and testing data is large.

DPL proposes a more efficient alternative: a single dynamic model that adjusts its processing depth based on the specific needs of the input image. This approach is guided by two primary factors: the visual quality of the image and the inherent difficulty of identifying the specific forgery.

Methodology: The "Drilling" Philosophy

The DPL framework is built around a recurrent architecture that progressively extracts features over multiple time steps. It consists of two main branches that guide this progression:

1. Visual Quality Guided Progressive Learning (VQPL)

This branch utilizes a Visual Quality Indicator (VQI) to estimate the level of compression in a video. The VQI leverages a pre-trained CLIP (Contrastive Language-Image Pre-training) model to categorize image quality into levels ranging from "perfect" to "bad." Based on this assessment, lower-quality videos are assigned more iterations (time steps) in a recurrent Gated Recurrent Unit (GRU) module, allowing the model more opportunities to find faint traces.

2. Forgery Identifiability Guided Progressive Learning (FIPL)

Recognizing that some forgeries are naturally more subtle than others regardless of compression, the Forgery Identifiability Indicator (FII) determines the effort required to spot specific manipulations. This ensures that even high-quality but highly realistic fakes receive the necessary computational attention.

Technical Innovations

  • Feature Selection Module (FSM): To prevent the model from redundant processing, the FSM uses a recurrent network to generate masks. These masks filter out forgery traces identified in early steps, forcing the model to focus on more "concealed" or difficult traces in subsequent iterations.
  • Two-Stage Training with PPO: Training a model to decide its own processing depth is a complex optimization problem. The authors employ Proximal Policy Optimization (PPO), a reinforcement learning algorithm, in the second stage of training. The model is rewarded based on the increase in detection confidence achieved through additional processing steps.
  • Weight Sharing: By using sequential units that share weights across time steps, DPL remains memory-efficient, avoiding the massive parameter counts associated with ensemble methods.

Experimental Evaluation

The authors validated DPL against several state-of-the-art (SOTA) methods using benchmark datasets including FaceForensics++ (FF++), CelebDFv2, FaceShifter (FSH), and FFIW10k.

Key Results

  • Superior Performance on Compressed Data: On the heavily compressed FF++ (c40) dataset, DPL achieved an 86.36% AUC, outperforming previous SOTA methods like QAD and UCF by approximately 1%.
  • Generalization and Robustness: DPL demonstrated strong cross-dataset generalization. When trained on one dataset and tested on entirely unseen ones (e.g., FFIW10k), it maintained a performance lead, suggesting it learns generalizable forgery features. Furthermore, the model showed high stability against common perturbations such as Gaussian noise and saturation changes.

Discussion and Limitations

While DPL offers a significant leap in cross-quality detection, it does introduce certain trade-offs. The primary limitation is computational overhead during inference; because low-quality videos require more time steps, they take longer to process than a standard single-pass model. Additionally, the system's effectiveness is tied to the accuracy of the CLIP-based indicators. If the VQI misjudges a low-quality video as high-quality, the model may not "drill" deep enough to identify the forgery.

Conclusion

DPL: Dual Progressive Learning represents a shift from static to dynamic DeepFake detection. By mimicking a "drilling" process guided by visual quality and forgery identifiability, the framework provides a robust solution for the challenges of real-world, compressed media. Future work may focus on more sophisticated fusion techniques for the dual branches to further enhance detection accuracy and efficiency.

Summary updated

BibTeX
@inproceedings{zhang2024dpl,
  author = {Dongliang Zhang and Yunfei Li and Jiaran Zhou and Yuezun Li},
  title = {{DPL: Cross-Quality DeepFake Detection via Dual Progressive Learning}},
  booktitle = {ACCV},
  year = {2024},
  doi = {10.1007/978-981-96-0960-4\_9},
}
Figure