Delayed Momentum Aggregation: Communication-efficient Byzantine-robust Federated Learning with Partial Participation
arXiv:2509.02970v3 Announce Type: replace Abstract: Partial participation is essential for communication-efficient federated learning at scale, yet existing Byzantine-robust methods typically assume full client participation. In the partial participation setting, a majority of the sampled clients may be Byzantine, once Byzantine clients dominate, existing methods break down immediately. We introduce delayed momentum aggregation, a principle where the central server aggregates cached momentum from non-sampled clients along with fresh momentum from sampled clients. This principle ensures Byzanti
The increasing scale and complexity of federated learning deployments necessitate more robust and efficient methods for handling unreliable participants, making research into Byzantine-robustness and partial participation critical.
This development proposes a solution to a key vulnerability in federated learning, potentially enabling more reliable and scalable AI systems, especially in scenarios with many distributed clients.
Existing federated learning methods are often vulnerable to Byzantine attacks when client participation is partial; this new approach offers a mechanism to mitigate that vulnerability, improving system resilience.
- · Federated Learning platforms
- · Edge AI developers
- · Critical infrastructure relying on distributed AI
- · Malicious actors targeting federated learning systems
- · Less robust federated learning frameworks
More secure and efficient federated learning deployments become feasible in real-world, large-scale applications.
Increased adoption of federated learning in sensitive domains like healthcare or finance due to improved security guarantees.
The proliferation of distributed AI systems could lead to new forms of collective intelligence that are resistant to manipulation.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.LG