5 June 2026 in IEEE Access
Federated Learning with Distillation (FD) has clients upload output logits evaluated on a shared public dataset, instead of gradients computed on private data. Although FD is often believed to offer stronger privacy protection than traditional federated learning, recent studies have shown that a malicious server can still extract sensitive information, such as reconstructing private samples or inferring label distributions in a private dataset [1], [2]. This suggests that FD is not immune to sensitive information leakages, raising concerns of other significant privacy threats such as property inference attacks. Motivated by these concerns, we explore property inference attacks (PIA) which aim to infer property population information (e.g., demographics or disease prevalence) of a private dataset. However, most existing PIA methods rely on the same underlying distribution as the victim’s private data, which is often unavailable in FD because the server only observes the public dataset distribution. This distribution shift can significantly degrade the effectiveness of existing PIA methods. To overcome this limitation, we introduce PI-FD, a new PIA method that employs synthetic data generation in combination with novel loss functions. Experiments on four datasets with 14 properties show PI-FD reaches 94.07% attack accuracy, highlighting the need to mitigate PIA threats.