Optimizing Pretrained Model with Squeeze-and-Excitation Blocks for Improved Image Forgery Detection

Detecting image forgeries is critical for ensuring the integrity of visual content which is very essential in the field of digital forensics. In this work, we investigate the use of pre-trained deep learning models in conjunction with sophisticated feature extraction methods, like squeeze-and-excitation (SE) blocks and hyperparameter fine-tuning, to provide notable gains in model performance. In order to detect forgeries, comparative analysis is performed on pre-trained models, notably ResNet50, VGG16, and EfficientNetB0, taking into account the Casia1.0 and Casia2.0 standard datasets. Performance is analyzed using metrics including accuracy, precision, recall, F1 score, and ROC curves, ensuring complete assessment of each model's efficacy. Results demonstrates our modified SE-enhanced EfficientNetB0 is effective in capturing intricate details that distinguish authentic images from forgeries, thereby advancing the capabilities of automated forgery detection systems.