Core Insights - The rapid development of AIGC technology has led to the generation of highly realistic content with simple prompts, but it also poses significant security risks such as fake news, identity fraud, and copyright infringement [1] - AI-generated image detection has become a fundamental security capability in the AIGC era, yet existing detectors perform well on benchmark datasets but struggle in real-world scenarios [1][3] - Tencent's Youtu Lab, in collaboration with research teams from East China University of Science and Technology and Peking University, has proposed the Dual Data Alignment (DDA) method to systematically suppress biased features and enhance the generalization ability of detectors across different models and data domains [1][18] Problem Identification - The root cause of detection issues lies in the construction of training data, where detectors rely on biased features rather than learning the essential characteristics that distinguish real from fake [3][4] - Systematic differences between real and AI-generated images lead to the learning of "shortcut strategies" by detection models, resulting in high accuracy on specific datasets but poor performance when faced with modified images [4] Proposed Solution - The DDA method aims to eliminate biases in training data through reconstruction and alignment, consisting of three main steps: pixel alignment, frequency alignment, and mixup [7][14] - Pixel alignment uses Variational Autoencoder (VAE) technology to reconstruct real images, ensuring consistency in content and resolution [8] - Frequency alignment addresses the loss of high-frequency information in JPEG-compressed real images, ensuring that the reconstructed images do not introduce new biases [9][12] - The final step involves mixing real and aligned generated images to enhance the alignment of true and false data [13] Experimental Results - The DDA method was evaluated under strict conditions, training a single universal model and testing it across various unknown and cross-domain datasets [15] - In a comprehensive test involving 11 different benchmarks, DDA outperformed in 10 of them, achieving a minimum accuracy (min-ACC) that was 27.5 percentage points higher than the second-best method [18] - The detection accuracy on the challenging "In-the-wild" dataset Chameleon reached 82.4%, demonstrating the model's effectiveness in real-world scenarios [18]