Accelerating surrogate model retraining in poisoning attacks against recommender systems

The researchers propose a Gradient Passing (GP) technique to accelerate surrogate model retraining for poisoning attacks against recommender systems. The method explicitly passes gradients between interacting user-item pairs during backpropagation to approximate cascading effects on the interaction graph. A single update can achieve an effect close to multiple original training iterations, bringing the surrogate model closer to the victim model under the same number of retraining rounds, thereby strengthening the poisoning attack, and the efficiency and effectiveness were validated on real datasets.