Federated learning is increasingly used to adapt a pretrained, centrally fine-tuned model to data that remain on user devices, rather than training a model from scratch. A new arXiv report examines the security implications of this approach, focusing on what can go wrong when the fine-tuning step itself is distributed across many clients.
The report identifies two main threats: privacy leakage, where attackers can infer information about user data, and data poisoning, where malicious participants can corrupt the model during fine-tuning. These risks are distinct from those in federated training from scratch, because the starting model already contains significant learned knowledge that can be exploited or altered.
The abstract is brief and does not detail specific experiments or countermeasures, so the full scope of the findings remains in the paper itself. Still, the report signals growing concern about the security of federated fine-tuning as a practical deployment pattern for pretrained models.