Poisoning attack on quality of service aware cloud API recommender system via surrogate generative adversarial network

To address the shortcomings of existing poisoning attack methods, where the generated fake user attack data suffers from poor attack effectiveness and high detectability, a poisoning attack method based on the surrogate generative adversarial network (S-GAN) was proposed. Firstly, K-means was used in the generative adversarial network to classify the data, and a self-attention mechanism was incorporated to learn the global features within each class, solving the problem of difficulty in capturing and mimicking key features of real users in sparse data, thereby enhancing the concealment of fake users. Secondly, a surrogate model was deployed to evaluate the attack effectiveness of the GAN-generated fake users, and the evaluation results were employed as a surrogate loss to optimize the GAN, thereby facilitating the attack effectiveness while considering the concealment of fake users. Experiments conducted on cloud API quality of service datasets demonstrate that the proposed method outperforms existing methods in balancing the effectiveness and concealment of attacks.