A Robust Content Identification System for Picture-In-Picture Attack Detection Using Trainable Background Removal and Perceptual Hashing Functions

This paper addresses copyright infringement in multimedia distribution with a model for identifying modified images, especially targeting PiP (Picture in Picture) attacks, which involve embedding and altering images in various ways. The model uses a three-step approach: first, detecting images of interest with a fine-tuned YOLOv5 model; second, employing the U2Net model from the rembg tool for background removal; and finally, applying perceptual hashing for compact hash code extraction. This methodology not only enhances security against unauthorized distribution but also aids in intellectual property protection. The model demonstrates high accuracy and reliability, with impressive F1 scores, making it an effective solution for copyright enforcement and digital media protection against complex manipulations.