Lightweight Convolutional Neural Network-Based Forgery Detection with Parallelized Feature Extraction

Being able to tell the difference between real and fake digital photos is an important part of multimedia forensics.This study proposes a hybrid architecture that integrates handcrafted and deep learning features to accurately detect tampering.The preprocessing pipeline uses block-wise Discrete Cosine Transform-Singular Value Decomposition (DCT-SVD) feature extraction, multi-quality Error Level Analysis (ELA), grayscale conversion, and Local Binary Patterns (LBP).A lightweight Convolutional Neural Network (CNN) with convolutional layers, batch normalization, global average pooling, and fully connected layers with softmax classification uses these features to make inputs with more than one channel.We used two large datasets, CASIA1 and CASIA2, to test the framework we made.We set aside 20% of the data for testing and 80% for training.Our system was right 98.13% of the time on CASIA1.It was also highly effective at finding the right things, with a precision of 98.09% and a recall of 98.17%.This meant that it was very well balanced between the two, with an F1-score of 98.12%.It still worked well, getting it right 96.68% of the time when we tested it on the bigger CASIA2 dataset.It was still highly effective at finding everything, with a recall rate of 97.49%, but it wasn't quite as precise, with a rate of 96.54%.We used several computers to get features at the same time and a special kind of computer chip to speed up the neural network part.Because of this, we could almost always work on problems as they came up.