Optimizing Data Efficient Image Transformers using K-fold Cross Validation for Passive Digital Image Splicing Forensics

The growth of advanced fake image editing tools makes detecting image splicing crucial due to its significant legal impact across various aspects of life. This research examines the effectiveness of the Data-efficient image Transformer (DeiT-Base) model in identifying manipulated images, highlighting its strengths in processing complex spliced images through advanced computational and contextual analysis coupled with the powerful K-fold cross-validation method. This integrated approach makes the most out of limited data to overcome overfitting, ensures rigorous model evaluation, and enhances the generalization capabilities of image splicing detection efforts. The model outperformed the existing technique on the benchmarked Columbia (Gray) dataset, achieving 100% precision and specificity and a 31% increase in F1-score. Other metrics also showed excellent results, suggesting the model could significantly improve image splicing detection. With this research’s exemplary outcomes, the technique sets a comprehensive approach for creating a sophisticated, data-efficient, and highly accurate image splicing detection model.