Diffusion-augmented adversarial poisoning in QoS-aware cloud API recommender systems

Data poisoning attacks pose a serious security threat to QoS-aware cloud API recommender systems by injecting carefully crafted fake user profiles into training data, thereby manipulating recommendation outcomes and compromising prediction reliability. Existing GAN-based poisoning methods often suffer from discriminator overfitting under limited interaction data, leading to pronounced distributional discrepancies between generated and real users, which significantly degrades both attack effectiveness and stealthiness. To address this limitation, we propose DA-GAN, a Diffusion-Augmented Generative Adversarial Network for poisoning attacks against QoS-aware cloud API recommendation systems. DA-GAN integrates diffusion probabilistic modeling into the adversarial training by constructing multi-step noise-augmented real users through a diffusion forward process, and generating fake users via noise-aligned posterior sampling in the reverse process. To further stabilize training, we design a multi-step noise-aligned discriminator that performs joint discrimination across multiple diffusion levels, providing reliable gradient supervision and effectively mitigating discriminator overfitting in small-sample settings. Extensive experiments on real-world cloud API QoS datasets demonstrate that DA-GAN consistently outperforms state-of-the-art poisoning baselines in both attack effectiveness and stealthiness across multiple mainstream QoS-aware recommender systems. Our findings reveal previously underestimated vulnerabilities in existing cloud API recommender systems and underscore the urgent need for diffusion-aware and attack-resilient defense mechanisms. Our code is available at https://github.com/Ethereal-Fairy/DAGANcode .