The CARECloud Project at the AAAI 2025 Conference: Securing Federated Learning
The CARECloud project participated in the 39th AAAI Annual Conference on Artificial Intelligence (AAAI-25) in Philadelphia in February 2025, highlighting the challenges of attribute inference attacks in federated regression tasks.
Federated learning allows multiple clients to collaborate to train a machine learning model while keeping their data local. However, this method is vulnerable to reconstruction attacks, where adversaries exploit exchanged messages to discover sensitive attributes. The CARECloud project presented new attribute inference attacks (AIAs) specifically designed for regression tasks. These AIAs consider scenarios where adversaries can eavesdrop on the exchanged messages or directly interfere with the training process.
The results show a significant increase in reconstruction accuracy, especially in heterogeneous client datasets. The effectiveness of these AIAs makes them better candidates for empirically quantifying privacy leakage in federated regression tasks.
By participating in this conference, the CARECloud project is demonstrating its commitment to securing federated learning and protecting data confidentiality. To find out more about the CARECloud project and its initiatives, visit their page.
