Federated Learning (FL) has revolutionized privacy-preserving AI by keeping data local and centralizing only the model updates. But what happens when relying on a central aggregator isn’t technically feasible, or when organizations refuse to grant a single entity the privileged role of “coordinator”?
In Talk 08 of ENSURE-6G Event #6, Carlos Segura from Telefonica Research introduced the paradigm of Peer-to-Peer (P2P) Learning. Presented in the context of the Go-Evolution project, the session explored how to successfully train machine learning models across a distributed network where there is no central server, no single global model, and total autonomy for every node.
The Paradigm Shift: Why Go Peer-to-Peer?
Standard FL relies on a star topology: all clients talk to one central server. While effective, this creates a single point of failure and requires total trust in the aggregator.
P2P Learning is essential for highly dynamic 6G edge networks where devices are unreliable, bandwidth is unpredictable, or organizations require direct autonomy over their collaborations. In a P2P architecture, the central server is completely removed. Instead, the intelligence and coordination responsibilities are pushed directly to the edge nodes.
The P2P Learning Loop
Without a server to dictate the rules, the learning loop transforms into a highly localized process:
- Local Training: A peer trains a model on its own private data.
- Neighbor Exchange: The peer exchanges its model updates only with a small, selected subset of neighbors (not the entire network).
- Local Assessment: The peer receives updates from its neighbors and independently evaluates them for freshness, compatibility, and trustworthiness.
- Local Aggregation: The peer combines the trusted information into its own model.
Crucially, there is no single identical global model at any point in the network. Each node develops a slightly different model tailored to its specific data distribution and the updates it chooses to accept from its neighbors.
The Complexity of the Graph
In centralized FL, the communication topology is simple. In P2P Learning, the network graph itself is a complex design parameter.
The graph determines how quickly information travels and how much bandwidth is consumed. If a node selects only similar neighbors, it reduces data diversity. If it connects to too many neighbors, it exhausts its bandwidth and needlessly expands its exposure surface to cyberattacks.
Interestingly, the graph topology acts as a natural privacy mechanism. Because local aggregation occurs at every “hop” in the network, individual contributions are continuously diluted. A node three hops away cannot infer the personal data of the original node—privacy becomes a function of graph distance.
Securing an Open Network: The SEC-FL Framework
Removing the central server also removes the central security authority. In an open P2P network, malicious nodes can easily attempt label flipping, gradient inversion, backdoor injections, or free-riding (benefiting from the network without contributing).
To combat this, the Telefonica team developed the SEC-FL P2P framework (built using Go and libp2p for the networking layer, and Python for the machine learning policy layer).
Because peers cannot rely on a server to ban bad actors, the SEC-FL framework equips every node with an experimental local security feedback loop.
- When a node receives an update, it evaluates the return similarity and reported learning loss.
- Over successive rounds, the node builds a reputation score for each of its neighbors.
- If a neighbor consistently sends suspicious or useless updates, the local node automatically reduces the weight of that neighbor’s contributions, switches to a robust aggregation method, or completely blocks the peer.
Conclusion
By decentralizing the coordination layer, P2P learning offers unparalleled autonomy and resilience for edge devices. As 6G networks grow increasingly dynamic, frameworks like SEC-FL prove that nodes don’t need a central master to collaborate safely and effectively—they simply need smart, locally enforced trust policies.
Watch the Full Talk:
