Recommendation algorithms can be manipulated not just to promote or bury products, but to make them systematically less fair — and a new paper shows exactly how.
Researchers published a method that uses reinforcement learning to inject fake user profiles into a recommender system in a way that amplifies existing bias. The approach combines a graph-based encoder to model how fake interactions relate to real ones, and a recurrent neural network to sequence the fake item injections. Crucially, it also includes a policy for choosing the gender of fake user profiles — designed specifically to undermine recommender systems that already use fairness-aware training. Tests ran against four recommendation model types on two real-world datasets.
Most adversarial research on recommender systems focuses on promotion or demotion attacks — getting a product ranked higher or lower. This work targets something harder to measure and easier to miss: whether the system treats different demographic groups equitably. If fairness audits become a regulatory requirement, as some jurisdictions are pushing toward, attacks like this one would let bad actors quietly degrade compliance without triggering obvious performance alarms.
The paper is framed as a defensive contribution — exposing a gap so it can be patched — but the gap it exposes is significant. A fairness-aware training pipeline is not a guarantee; it is a surface that now has a documented exploit.