<p>The spread of forged images can easily and often be performed in social media which is then compressed into the JPEG format. The purpose of such compression is to reduce transmission bandwidth or storage space. Deployment of compressed images for forensic attacks to automatically classify indications of image forgery. This is because the classification system uses image learning data, not all of which have been compressed. For learning the classification of counterfeit images on social networks, the system needs to compress the fake images or original images in the learning data to JPEG format. Deep learning can be used to identify image forgery, but it requires significant processing time and resources. This study investigates lightweight deep learning, specifically the ShuffleNet V2 architecture, for detecting forged images that have undergone JPEG compression under multiple recompression scenarios typical of social media platforms. Then feature extraction is accelerated with the ShuffleNet V2 modification by adding the Ghost module as a replacement for the last convolution (Conv5) and the Squeeze-and-Excitation (SE) module. Modifications were also made by replacing the activation function using the FReLU activation funnel. The research results that combined several datasets with images compressed in JPEG format using various compression variations indicate that the modified model on the ShuffleNet V2 layer performs better compared to several other lightweight deep learning models. These include the basic architecture of ShuffleNet V2, MobileNetV2, and ResNet-50, as seen in terms of accuracy, precision, recall, and F1-score values.</p>