This paper introduces AdFlush, a new machine learning model designed to combat web content manipulation. Through evaluating 883 features, 27 key features were selected for optimal performance. Testing on 10,000 real-world websites resulted in an impressive F1 score of 0.98, outperforming other models like AdGraph, WebGraph, and WTAgraph. AdFlush also reduces computational overhead by 56% CPU and 80% memory compared to AdGraph. It showed superior resilience against adversarial manipulations with F1 scores ranging from 0.89 to 0.98. The model maintained a high F1 score above 0.97 without retraining over six months, proving its effectiveness in detecting malicious web content.
https://dl.acm.org/doi/abs/10.1145/3589334.3645698