Performance Assessment of Fine-Tuned VGG16 Model on the MDRF Retouched Face Dataset

Background: Objective: Despite being widely employed on digital platforms, facial retouching frequently causes issues since it compromises the truthfulness and trustworthiness of visual data and the present work is focused on it. Method: The classification of retouched face photos using a transfer learning approach is thoroughly examined in this research by fine tuning the VGG16 model pre-trained on “ImageNet” weight. The specific Caucasian ethnicity is used, which includes both actual and retouched face samples, and different intensities of retouching are applied to each sample separately. Finding: The experimental results show that the proposed transfer learning approach and fine-tuning of the existing model can achieve better classification efficiency with training on a smaller dataset. A 60 % - 40 % train-test split ratio outperforms other split ratios mentioned in the paper with 98.12% precision, recall, and F1 score in the classification of retouched face samples. The fine-tuned VGG16 model with Adam optimizer gives an overall accuracy of 97.50 % as compared to existing works. Novelty: For model training, the Adam optimizer is used, which adaptively modifies the network weights and dramatically lowers training and validation loss, improving classification performance overall. Keywords: Transfer Learning, VGG16, Fine-tuning, MDRF, Adam