Understanding and Improving Adversarial Collaborative Filtering (ACF) for Robust Recommendation

The paper proposes Personalized Magnitude Adversarial Collaborative Filtering (PamaCF), which applies personalized perturbation magnitudes to different users according to their embedding scales, significantly improving recommendation performance while defending against data poisoning attacks. Theoretical analysis proves that, under the same number of training epochs, ACF achieves lower recommendation error than traditional collaborative filtering on both clean and poisoned data. Extensive experiments show that PamaCF can effectively defend against multiple types of poisoning attacks.