<![CDATA[The rapid advancement of generative artificial intelligence models has escalated the risks of deepfake image dissemination, necessitating a paradigm shift from binary detection to provenance attribution, i.e., identifying the specific AI source model. Although various Convolutional Neural Network (CNN) architectures have proven highly accurate in classifying image sources under ideal conditions, real-world implementations frequently encounter spatial degradations, such as blurring effects, which destroy AI-generated high-frequency artifacts. This study presents a comparative analysis of the robustness of ResNet50 and EfficientNetB0 architectures in attributing multi-class deepfake images (DALL-E, Midjourney, Stable Diffusion, and Real Images) distorted by Gaussian Blur attacks, using a transfer learning approach to evaluate the degradation rate of both models. Experimental results indicate that while EfficientNetB0 achieved an accuracy of 93.89% under ideal conditions, its performance dropped significantly by 6.56% to 87.33% when images were blurred, due to the sensitivity of its depthwise convolution layers. In contrast, ResNet50, which achieved 95.70% accuracy under ideal conditions, demonstrated far superior robustness with only a 1.58% accuracy degradation, retaining 94.12% accuracy after the Gaussian Blur attack. This advantage is empirically validated as the effect of skip connections in the residual network, which preserve the spatial integrity of semantic feature representations. This study recommends the use of residual-based architectures for digital forensic systems operating in real-world environments that are susceptible to compression and degradation.]]>
Copyrights © 2026
