TabLeak: Tabular Data Leakage in Federated Learning
This work addresses a significant privacy risk for users of federated learning systems in sensitive domains, revealing vulnerabilities that were previously unexplored for tabular data.
The paper tackles the problem of data leakage in federated learning for tabular data, which is critical in high-stakes applications like healthcare and finance, and shows that their TabLeak attack successfully extracts private data with over 90% accuracy in settings previously considered safe.
While federated learning (FL) promises to preserve privacy, recent works in the image and text domains have shown that training updates leak private client data. However, most high-stakes applications of FL (e.g., in healthcare and finance) use tabular data, where the risk of data leakage has not yet been explored. A successful attack for tabular data must address two key challenges unique to the domain: (i) obtaining a solution to a high-variance mixed discrete-continuous optimization problem, and (ii) enabling human assessment of the reconstruction as unlike for image and text data, direct human inspection is not possible. In this work we address these challenges and propose TabLeak, the first comprehensive reconstruction attack on tabular data. TabLeak is based on two key contributions: (i) a method which leverages a softmax relaxation and pooled ensembling to solve the optimization problem, and (ii) an entropy-based uncertainty quantification scheme to enable human assessment. We evaluate TabLeak on four tabular datasets for both FedSGD and FedAvg training protocols, and show that it successfully breaks several settings previously deemed safe. For instance, we extract large subsets of private data at >90% accuracy even at the large batch size of 128. Our findings demonstrate that current high-stakes tabular FL is excessively vulnerable to leakage attacks.