Defeating Anti Ad Blockers

Abstract

Perceptual ad blockers came with the claim that they finished the arm race between ad blockers and anti-ad blockers. They used the technique to visually identify advertisement instead of blocking domains. But adversarial attacks successfully defeated all the classifiers that were using to classify ads. But the Classifier that we purposed is not only detecting Ad Choice logo in advertisement images with 99% accuracy but also undefeated by all adversarial attacks that has been performed to bypass perceptual Ad blockers. In our research we took dataset of 1000 advertisement images with ad choice logos and detect ad choice logo with different object detection models. We found cross correlation classifier with 99 percent detection accuracy. Next task was to evaluate this classifier against different adversarial attacks. We performed all adversarial attacks that “Ad Versatile: Perceptual Ad Blocking meets Adversarial Machine Learning” paper performed to defeat perceptual ad blockers. We also used some other scale base noises but our classifier successfully detects ad choice logo in all noisy images. Machine learning classifiers are easy to defeat using adversarial noise but simple computer vision algorithms are less prone to adversarial attack. So our work proves that it is not easy to defeat perceptual adblockers as our classifier can extract all images from webpage and by detecting ad choice logo it can classify it as an advertisement.