Generative recommendation models, which rely on autoregressive decoding, are increasingly being used as an alternative to classical ranking-based systems. A new paper on arXiv introduces AlignPoison, a method designed to perform conditional targeted poisoning of these models via decoding-aligned prefixes.
The abstract notes that while the modeling flexibility of generative recommenders has been widely studied, their resilience to data poisoning has received less attention. AlignPoison appears to address this gap by demonstrating a specific attack vector, though the full details of the method and results are not included in the abstract.
This work highlights a potential security concern for systems that adopt generative approaches for recommendation, suggesting that data poisoning remains a relevant threat even for more flexible models.