Improving the weakest link: vulnerability-aware adversarial training (VAT) for robust recommender systems

Against poisoning attacks that inject fake users into recommender system training data, researchers proposed vulnerability-aware adversarial training (VAT). The study found that attacks usually affect only some vulnerable users, and that the more closely the system fits a user's training data, the more susceptible that user is to attack influence; VAT accordingly applies perturbations of adaptive magnitude to each user. Experiments show that VAT reduces attack success rates across multiple recommender models and attack types while maintaining and possibly even improving recommendation quality.