260115 Enhance publishing|One-step Reconstruction Diffusion Model-based Poisoning Attack onQoS-aware Cloud API Recommender Systems

The Quality of Service (QoS) aware cloud application programming interface (API) recommendation system effectively alleviates the information overload challenge caused by the continuous growth of cloud APIs by guiding users to discover high-quality cloud APIs. However, existing research on QoS aware cloud API recommendation systems mainly focuses on improving recommendation accuracy, ignoring the security risks brought by poisoning attacks.

Therefore, this study proposes a preference guided poisoning attack framework (PDPA) based on a one-step reconstruction diffusion model from the perspective of attack learning defense to simulate poisoning attacks and reveal the vulnerability of cloud API recommendation systems. Firstly, PDPA uses a one-step reconstruction diffusion model to separately model the QoS and call distribution of real users regarding cloud APIs, generating false user QoS and call behavior similar to real users.

Next, PDPA selects fake users who have a preference for calling the target cloud API to simulate poisoning attacks, effectively reducing the interference of the target cloud API on the concealment of fake users and ensuring the effectiveness of fake user attacks. Finally, extensive experiments were conducted on real-world datasets, and the results demonstrated the vulnerability of QoS aware cloud API recommendation systems under poisoning attacks, as well as the superior attack effectiveness and concealment of false users generated by PDPA compared to baseline methods.