Image Forensics of Compressed Image on Social Media with Lightweight Deep Learning

The widespread use of forged images on social media often involves compressing them into the JPEG format to save bandwidth or storage. For automatic detection of these forgeries, a classification system is used. However, this system is designed based on image data that is not always compressed. To effectively classify forgeries on social media, the learning data should include both fake and original images compressed to the JPEG format. Deep learning offers a solution to detect image forgery, but its high processing demands are a challenge. This study employs the lightweight ShuffleNet v2 architecture to mitigate this, further optimizing feature extraction by substituting the activation function with the FReLU activation funnel. Our preliminary research utilized JPEG-compressed images to create a relevant dataset. The test aims to gauge the efficacy of the modified ShuffleNet v2 in spotting fake images compressed via the Facebook application.