Enhancing Passive Digital Image Splicing Forensics Using Data-Efficient Image Transformer and Lightweight Pretrained ShuffleNet-V2 Models - A Comparative Analysis
Chithra Raj N., Maitreyee Dutta, Jagriti Saini · 2024
An investigation into the challenge of detecting image splicing, a prevalent form of digital image manipulation, using advanced deep learning techniques is of paramount importance, owing to its serious legal implications in every realm of life. The study explores two models focused on passive image splicing forensics: the lightweight, pretrained ShuffleNet-V2 and the Data-efficient Image Transformer (DeiT-Base). ShuffleNet-V2, known for its efficiency and speed, demonstrates a significant reduction of 3% in trainable parameters compared to the state-of-the-art method, making it ideal for real-time applications. DeiT-Base excels in complex image processing due to its superior computational and contextual analysis capabilities. Both models were tested using the Columbia dataset and showcased remarkable accuracy, with ShuffleNet-V2 achieving 94.44% and DeiT-Base 97.22%. They outperformed existing techniques, such as MobilePSP-Net, TransU2-Net by delivering 100% precision and specificity. The F1-score (0.9412 and 0.9714, respectively) and recall (88.89% and 94.44% respectively) also showed notable improvements, indicating these models' potential to enhance image splicing detection methods. With this research's excellent outcomes, both models hold considerable promise in performance, as their distinct approaches provide valuable perspectives for advancing image splicing detection methodologies.