As digital advertising continues to expand, recommendation systems face two major challenges: recommendation distortion due to fake user attacks and limited targeting precision caused by data sparsity. Traditional methods rely on manual rules, making them ineffective against complex attack patterns. This project focuses on defending against sybil attacks in collaborative filtering scenarios by building an intelligent detection system. Using convolutional neural networks, the system learns deep user behavior features to accurately identify forged ratings and abnormal interactions. The project innovatively proposes a dynamic feature extraction framework that automatically learns local and global correlations in user ratings, overcoming the limitations of manual feature engineering. The results show that the model achieves an F-value of 0.89 in detecting hybrid attacks on the MovieLens and Netflix datasets, a 23.6% improvement over traditional PCA methods. Coverage of long-tail attacks is also significantly enhanced. This work constructs a robust security foundation for personalized recommendation systems, promoting the healthy development of the digital marketing ecosystem.

网络结构

用户画像模型构建

个性化广告推荐