ITML’s federated learning (FL) framework, developed within the CyberSecDome project, provides a powerful solution for privacy-aware information and knowledge sharing across cybersecurity infrastructures. Designed to facilitate collaboration among diverse organizations without compromising sensitive data, this framework enables the secure exchange of AI model parameters — such as weights, hyperparameters, or evaluations —without revealing the raw data used in training. It offers a practical and privacy-focused response to the challenge of sharing intelligence across silos in an increasingly connected digital threat landscape.

At the heart of the solution is Flower, an open-source FL framework that ensures scalability, flexibility, and robustness. Flower allows the orchestration of federated learning across a variety of client types — ranging from servers and smartphones to embedded systems — while supporting all major machine learning frameworks such as PyTorch, TensorFlow, and XGBoost. Custom aggregation strategies, secure communications (e.g. gRPC), and adaptive network topologies make this solution suitable for real-world, large-scale deployment. Importantly, Flower’s compatibility with asynchronous learning and heterogeneous clients ensures resilience even under varying connectivity and compute conditions.

The framework integrates a suite of advanced privacy-enhancing technologies (PETs) to guard against data leakage and inference attacks. These include:

• Differential Privacy (via OpenDP, PyDP, Privacy on Beam),
• Homomorphic Encryption (via TenSEAL), and
• Secure Multiparty Computation (SMC).

By adopting these techniques, the system not only ensures strong data protection guarantees but also meets compliance requirements such as GDPR. Furthermore, the inclusion of Presidio and pyCANON supports fine grained PII detection and dataset anonymization, enabling organizations to assess and mitigate re-identification risks.

By enabling decentralized training, ITML’s FL framework empowers organizations to continuously refine shared cybersecurity models while retaining full control over their private data. This approach supports broad-spectrum threat detection, real-time adaptation, and a more robust collective defense posture. Whether deployed in critical infrastructure, enterprise environments, or public-sector systems, the framework is engineered for trustworthy AI collaboration in an era of evolving digital threats.