基于全局频谱感知网络的深度伪造检测OA

The rapid development of Generative Adversarial Networks(GANs)led to each generative model introducing unique artifacts.Current forgery detection models performed poorly when faced with images generated by multiple models.To address this,a Global Spectral Awareness Network(GSANet)was proposed.A Multi-Scale Spectral Learning module(MSSL)was designed,which directly performed multi-scale learning on the spectrum to enhance the ability to extract frequency domain features.High-Frequency Information Extraction module(HFIE)was employed,which forced the model to consistently focus on high-frequency information.The tendency of overfitting was effectively mitigated through emphasizing the importance of high-frequency information.Dilated Squeeze Attention module(DSA)was constructed.Dilated convolution was combined with attention mechanisms to replace the 3×3 convolution in the ResNet bottleneck block,and a DSA Block was formed for global feature learning.This enabled the model to achieve more generalized deepfake detection.Validation on eight different GAN-generated test sets demonstrated a significant 4％performance improvement.