Multi-scale perturbation fusion adversarial attack on MTCNN face detection system

Researchers have made significant progress in fake face image detection in recent years, which can accurately identify fake face images, but it is still challenging to prevent malicious face-swapping. By analyzing the popular face swap program faceswap, we found that adding perturbations to images to disable face detectors is an effective way to avert face swaps from happening. We propose an adversarial attack method that performs multi-scale interpolating perturbations for the MTCNN face detection system. This method can improve the stability of adversarial samples, resist the perturbation failure caused by scale changes, and achieve attack effects on single and multi-person images while maintaining the cleanliness of images so that images can continue to upload to social media platforms.