<![CDATA[The rapid advancement of deepfake technology poses a serious threat to digital security. Unfortunately, the majority of existing detection systems still rely on complex network architectures with heavy and costly computational requirements. Therefore, this study evaluates the reliability of MesoNet-4 a lighter Convolutional Neural Network (CNN) architecture and efficiently detecting deepfakes. The evaluation was conducted using a dataset of 140,000 faces from Kaggle, simulated in the Google Colab environment. Experimental results show that the model achieved a high accuracy rate of 94.00% when tested on data from the same distribution (in-dataset). However, this performance dropped significantly to 50.00% when subjected to out-of-dataset testing. In this scenario, the model experienced a complete detection failure by classifying all fake image samples as real images. This phenomenon indicates that lightweight architectures are highly susceptible to overfitting to the specific photographic characteristics of the training data, thereby failing to adapt to domain shifts in real-world conditions. This study concludes that high accuracy in a controlled environment does not guarantee the model’s practical reliability. Moving forward, the development of detection systems needs to prioritize cross-domain generalization capabilities and integrate dynamic preprocessing modules, such as automatic face detection.]]>
Copyrights © 2026
