<![CDATA[Facial biometric authentication systems are increasingly vulnerable to sophisticated digital manipulations, particularly deepfakes. Conventional single Convolutional Neural Network (CNN) models often encounter overfitting issues, especially when trained on high-resolution datasets containing subtle manipulation artifacts. To address this issue, this study evaluates and compares the effectiveness of applying L1 Regularization and L2 Regularization parameters to the final classification layer of a fine-tuned ResNet-50 architecture. The models were comparatively evaluated using the public CIPLAB Real and Fake Face Detection dataset. To ensure an objective stability measurement, a multi-seed evaluation method was applied using three random initializations. Experimental results indicate that the baseline model without regularization exhibited a high level of prediction fluctuation, with an accuracy standard deviation of 0.0387. The application of L1 Regularization achieved an average testing accuracy of 68.78%, but failed to resolve the stability issue, as it still recorded a high standard deviation of 0.0380. Conversely, L2 Regularization was proven to significantly mitigate the risk of overfitting, emerging as the most superior approach. This is evidenced by a sharp decrease in the accuracy standard deviation to 0.0105, while securing a consistent average testing accuracy of 68.78%. These findings emphasize that applying an L2 Regularization penalty to a single CNN model offers a far more stable, effective, and consistent solution for detecting facial image manipulations compared to both L1 Regularization and conventional models.]]>
Copyrights © 2026
