Abstract
Deepfake detection models often suffer from performance degradation when deployed on Online Social Networks (OSNs) due to heavy JPEG compression. This compression introduces "block effects" that mimic the checkerboard artifacts characteristic of generative models, leading to high false-positive rates. We present PLADA (Pay Less Attention to Deceptive Artifacts), a framework designed to enhance detection robustness by actively suppressing compression-related features. PLADA employs a dual-stage attention mechanism, the Block Effect Eraser (B2E), and an Open Data Aggregation (ODA) module to focus on genuine forgery signatures. Experimental results across 26 datasets demonstrate that PLADA significantly outperforms existing state-of-the-art methods in compressed environments while maintaining high accuracy on raw images.
Introduction
While deepfake detection has seen significant progress in controlled, high-resolution environments, real-world application on platforms like Facebook, X, or WeChat remains challenging. Images shared on these OSNs undergo aggressive JPEG compression, which divides images into 8x8 pixel grids. This process creates structured "block effects" that standard neural networks often confuse with the artifacts produced by Generative Adversarial Networks (GANs).
As shown in the figure above, the distinction between a real compressed image and a deepfake becomes blurred as compression noise masks the underlying generative traces. Current detectors often "pay too much attention" to these deceptive artifacts, failing to generalize to the varied compression levels found online.
Methodology
The PLADA framework addresses the challenge of deceptive artifacts through two primary components:
1. Block Effect Eraser (B2E) Module
The B2E module reconfigures the model's attention to ignore 8x8 block boundaries and focus on global inconsistencies.
- Residual Guidance (RG): Utilizes a "Guide Prompt Pool" containing prior knowledge of compression patterns, allowing the backbone to subtract these features from the representation.
- Coordination Guidance (CG): Refines attention at both fine-grained and coarse-grained levels to ensure sensitivity to facial inconsistencies rather than grid noise.
- Adversarial Learning: Employs a compression-discard branch with a Gradient Reversal Layer (GRL). This setup penalizes the model for identifying whether an image is compressed, forcing it to learn features invariant to JPEG noise.
2. Open Data Aggregation (ODA) Module
To overcome the scarcity of paired (raw and compressed) training data, the ODA module uses a sophisticated loss function. It clusters images into four categories (Real/Fake and Raw/Compressed) and calculates their feature centers. The module maximizes the distance between "Real" and "Fake" clusters while minimizing the distance between raw and compressed versions of the same class, effectively utilizing unpaired data to build robustness.
Evaluation and Results
The authors evaluated PLADA across a comprehensive suite of 26 datasets, including 17 GAN-based subsets (e.g., ProGAN, StyleGAN) and newer Diffusion Model (DM) generators like DALL-E and Stable Diffusion.
- Robustness to Compression: PLADA achieved a mean accuracy of 76.7% across GAN datasets, outperforming previous state-of-the-art methods like ODDN and QAD.
- Data Efficiency: In tests where only 10% of training data was paired, PLADA maintained a 7% performance lead over competitors, highlighting the effectiveness of the ODA module.
- Performance on Raw Data: Unlike many robust models that sacrifice "clean room" performance, PLADA maintained a high Average Precision (AP) of 80.4% on raw images.
Conclusion
PLADA represents a significant advancement in making deepfake detection viable for real-world social media environments. By shifting the focus from deceptive compression artifacts to intrinsic forgery cues, the framework achieves superior robustness without sacrificing accuracy on high-quality content. Future research will explore extending this approach to other OSN-specific artifacts, such as resizing and beautification filters, and integrating multi-modal signals from Large Language Models to further refine detection accuracy.
Summary updated