Deliverable 2.15: Communication protocols

Deliverable 2.15 on “Communication protocols” reports on the progress within Task 2.6 regarding communication protocols for data- and model-centric federated methods. In this deliverable, we report on how the Flower federated learning framework communicates during federated learning experiments and specifically how this communication can be secured. Specifically, we cover the following methods to secure communication:

• Secure Aggregation
• Differential Privacy
• Trusted execution environments

Secure Aggregation and Differential Privacy have been validated and disseminated during Project Workshop 5, organised virtually on October 1, 2025.

Based on the code provided to participants in workshop 5, we have concluded that the decrease in model performance is negligible for both Secure Aggregation and Differential Privacy. The runtime remains constant when utilizing Differential Privacy, whereas Secure Aggregation adds about 13% runtime due to the additional communication required between the server and clients. In addition, the operational overhead of deploying either technique is minimal when utilizing the Flower federated learning framework, although Differential Privacy requires model and dataset-specific tuning of the hyperparameters. Finally, we have concluded that Trusted Execution Environments can provide additional security guarantees for the global model, but require significant technical setup on both the server and all clients, which is highly dependent on the available hardware, making it impractical to deploy compared to Secure Aggregation and Differential Privacy.