CAFE: Catastrophic Data Leakage in Vertical Federated Learning

AI summaryThe researchers propose CAFE, a data leakage attack against vertical federated learning that can efficiently recover batch training data from shared aggregated gradients and has a theoretical basis. The paper points out that increasing the batch size is often regarded as a defense strategy against leakage, whereas in vertical federated learning experiments CAFE can recover large-batch data with higher quality than existing attacks, indicating that this defense approach is insufficient to protect privacy. The authors also provide a practical countermeasure to mitigate CAFE, and argue that privacy data involved in training in standard federated learning, especially in the vertical setting, faces a high risk of gradient leakage. Code is available at https://github.com/DeRafael/CAFE。