Journal of Basic and Applied Research International

Advanced image-forgery methods in the contemporary digital environment threaten media integrity and information security. Although convolutional neural networks (CNNs) are well suited to extracting semantic information from data, they are less effective at identifying subtle forensic information, particularly when combined with a traditional multilayer perceptron (MLP) classification head. This paper proposes a new hybrid architecture, Neuro-Forensics, which combines a ResNet-50 feature-extraction backbone with a Kolmogorov–Arnold network (KAN) classification head to improve non-linear function approximation. A multimodal feature-fusion approach is used to form a six-channel input array comprising spatial RGB information and computed features, including error level analysis (ELA), high-frequency noise residuals, and discrete cosine transform (DCT) coefficients. An ablation study was conducted on 22,614 selected images from the CASIA and MICC-F2000 benchmarks using a random train–validation–test split. The proposed ResNet-KAN hybrid increased overall test accuracy from 78.08% for the six-channel Res-CNN comparator to 78.23% and increased correct splicing classifications by 8.2% relative to the four-channel baseline. The model achieved areas under the curve (AUCs) of 0.98 for retouching and 0.94 for splicing. Grad-CAM provided visual interpretations of the regions influencing the model's decisions, although pixel-level localisation metrics were not reported. The findings suggest that learnable polynomial activation functions on network edges may approximate high-frequency forensic signals more effectively than conventional linear layers. However, robust cross-dataset evaluation is required to address the risk of data leakage.