Decentralized Federated Learning

Nodes exchange model updates directly with randomly selected neighbors using gossip communication rather than routing through a central server. This epidemic-style dissemination ensures that updates eventually propagate to the entire network with logarithmic convergence time.

• Relies on push-pull averaging to blend local models
• Tolerates node churn and intermittent connectivity
• Example: A consortium of 50 banks training a fraud detection model without any single institution coordinating the round

Model updates are validated and ordered using distributed ledger technology, providing an immutable audit trail of contributions. Smart contracts can automate reputation scoring and incentive distribution based on update quality.

• Prevents free-riding where nodes benefit without contributing
• Enables tokenized reward mechanisms for honest participation
• Example: A decentralized AI marketplace where contributors earn tokens for providing high-quality gradient updates verified by on-chain consensus

Decentralized topologies must withstand adversarial nodes that send corrupted or manipulative updates. Robust aggregation rules like Krum, median-based, or trimmed-mean operators filter out anomalous contributions before model blending.

• Defends against model poisoning attacks
• Maintains convergence even with up to 33% malicious nodes
• Example: A defense coalition training on classified sensor data where compromised nodes may attempt to skew the global model

Nodes apply local differential privacy by clipping gradients and injecting calibrated Gaussian noise before sharing updates with peers. This provides plausible deniability for individual training records without trusting any aggregator.

• Privacy budget (ε) is consumed locally per round
• Eliminates the need for a trusted curator
• Example: Smartphone keyboards learning typing patterns where each device adds noise to its update before gossiping, preventing reconstruction of any user's messages

The network graph is continuously reorganized based on node availability, bandwidth, and computational capacity. Lightweight discovery protocols allow nodes to join or leave without reinitializing the training process.

• Supports heterogeneous hardware from GPUs to microcontrollers
• Adapts routing to minimize communication diameter
• Example: An IoT fleet of 10,000 sensors where devices self-organize into sub-graphs based on geographic proximity and available power

Nodes collaboratively compute the average of their model updates using secret sharing or masking protocols so that no single peer can inspect another's raw contribution. Only the aggregated result is revealed to participants.

• Uses Shamir's Secret Sharing or additive masks
• Provides cryptographic privacy even against colluding nodes
• Example: Three competing pharmaceutical companies jointly training a drug discovery model where each learns only the averaged gradient, never the proprietary data or individual updates of rivals

A fully decentralized protocol where nodes exchange and average models directly with randomly selected peers. Gossip Learning eliminates the single point of failure inherent in centralized federated topologies by using epidemic-style communication. Each node periodically selects a neighbor, sends its current model, and updates its local copy upon receiving a peer's model. This approach provides robustness to node churn and network partitions, making it ideal for highly dynamic edge environments. Convergence is achieved through iterative pairwise averaging, though communication overhead can be higher than server-coordinated methods.

A mechanism that uses distributed ledger technology to coordinate model updates without a trusted central party. Nodes submit local gradient updates as transactions, and a consensus protocol (e.g., Proof-of-Stake or Practical Byzantine Fault Tolerance) validates and commits an aggregated global model to the chain. This provides an immutable audit trail of all contributions and prevents malicious nodes from poisoning the collaborative model. Smart contracts can automate incentive distribution, rewarding participants proportionally to the quality of their updates. Frameworks like OpenFL and FederatedScope support blockchain backends.

A decentralized aggregation strategy where nodes collaboratively compute a global model through direct neighbor exchanges rather than relying on a central server. Techniques include all-reduce averaging, where nodes collectively sum and distribute model parameters, and ring-based aggregation, where updates traverse a logical ring and are accumulated hop-by-hop. This architecture eliminates the communication bottleneck at the server and reduces vulnerability to single-point censorship. It is particularly effective in permissioned networks where all participants are known and trusted, such as a consortium of hospitals training a diagnostic model.

The application of formal privacy guarantees in peer-to-peer federated systems where there is no trusted curator to manage the privacy budget. Each node independently adds calibrated Gaussian or Laplacian noise to its model updates before sharing them with neighbors. This ensures that even if an adversary observes all messages on the network, they cannot infer the presence of any single training record. The challenge lies in coordinating the privacy budget (ε, δ) across a fully distributed topology without a central accountant. Techniques like secure aggregation combined with local differential privacy provide strong end-to-end guarantees.

Defensive algorithms designed to protect decentralized federated learning from adversarial nodes that submit arbitrary or malicious model updates. Without a central server to vet contributions, each node must independently apply robust aggregation rules. Methods include: Krum, which selects the update closest to a majority of its neighbors; Trimmed Mean, which discards extreme values before averaging; and Median-based aggregation, which uses coordinate-wise medians to neutralize outliers. These techniques are critical for maintaining model integrity in open, permissionless networks where Sybil attacks are possible.

A cryptographic protocol that allows a group of mutually distrusting nodes to jointly compute an aggregated model without revealing their individual updates to any single party. Using secret sharing, each node splits its gradient into fragments and distributes them to peers. The peers collaboratively compute the sum using secure addition circuits, and the result is reconstructed only when a threshold of honest nodes participates. This provides information-theoretic privacy against collusion and is often paired with differential privacy for defense-in-depth in highly sensitive applications like cross-border financial modeling.