Enhancing Image Forgery Detection through Attention Mechanisms and Knowledge Distillation

Abstract

: Image editing tools are common and easy to use which has made image tampering a concern in journalism, law, and online media. Convolutional Neural Networks (CNNs), have improved forgery detection. However, single CNN models still find it hard to notice minute changes such as texture differences or compression marks. The present study explores improved CNN models for image-tampering detection. Three main models: ResNet-18, MobileNetV3-Small, and EfficientNet-B0 were tested. EfficientNet-B0 gave the best starting accuracy. Two methods were applied to make the models perform better. At first, the Convolutional Block Attention Module (CBAM) was added to ResNet-18 so the model could better attend to key parts of an image. Next, knowledge distillation helped a smaller MobileNetV3-Small model learn from a larger network. The ResNet-18 + CBAM model achieved 0.844 accuracy surpassing the baseline model while the distilled MobileNetV3-Small retained 91% of its teacher model accuracy with ~15 ms/image inference which is about 10 times faster. These changes improved accuracy and stability while keeping the network fast and lightweight.