TRUSTINCloudS at PETS 2025: Innovations in Cybersecurity and Federated Learning
From July 14 to 19, 2025, Washington DC will host the 25th edition of the Privacy Enhancing Technologies Symposium (PETS). This event brings together global experts to discuss advances in privacy technologies. The TRUSTINCloudS project, which develops solutions for cybersecurity challenges in Cloud environments, will be presenting its innovative work.
A federated learning environment that respects privacy and promotes equity
Federated learning enables distributed model training between multiple data owners, orchestrated by a central server. However, this approach has shortcomings in terms of confidentiality and fairness. Model updates can lead to membership inference attacks, while unbalanced local datasets can introduce classification biases, particularly for minority groups.
The TRUSTINCloudS project has developed a new inference attack exploiting fairness metrics when training the global model. To counter this threat, they propose a fairness-sensitive encrypted domain aggregation algorithm that is differentially private by design. This uses the approximate loss of precision of the homomorphic multi-key CKKS threshold encryption scheme.
The results, presented at PETS 2025, demonstrate the effectiveness of this approach in terms of fairness and confidentiality, validated by experiments on three real data sets.
To find out more, read the full paper here.
