Federated Learning

The core architectural shift where a global model is distributed to edge devices or local servers. Each node trains the model on its local dataset, and only the model updates (gradients or weights) are sent back to a central coordinating server. The raw data never leaves the device of origin.

• Data Locality: Sensitive information remains siloed on the generating device.
• Reduced Data Transfer: Only mathematical vectors are communicated, not terabytes of raw logs.
• Heterogeneous Nodes: Supports training across devices with vastly different compute capabilities and network availability.

Federated learning alone does not guarantee privacy; model updates can leak information. Differential Privacy (DP) is mathematically integrated by clipping per-client updates and adding calibrated Gaussian noise to the aggregated model. This provides a provable bound on the privacy loss, ensuring an adversary cannot determine if a specific individual's data was included in the training set.

• Epsilon Budget: A parameter controlling the privacy-utility trade-off.
• DP-FedAvg: A variant of the standard algorithm that ensures the global model satisfies differential privacy guarantees.

A cryptographic technique that ensures the central server can only decrypt the sum of all client updates, not any individual client's contribution. Using Secure Multi-Party Computation (SMPC) or Homomorphic Encryption (HE), the server computes the aggregate model while the individual updates remain encrypted in transit and during computation.

• Zero-Knowledge Server: The coordinator remains blind to individual data contributions.
• Masking Vectors: Clients use pairwise secrets to mask their updates, which cancel out only in the final sum.

Two distinct operational paradigms exist. Cross-silo federated learning involves a small number of reliable, stateful clients, typically organizations like hospitals or banks, with large compute resources. Cross-device involves millions of unreliable, stateless edge devices like smartphones or IoT sensors with limited bandwidth and battery life.

• Cross-Silo: Focus on vertical scaling and complex model architectures.
• Cross-Device: Focus on fault tolerance, straggler mitigation, and extreme compression.

A specialized paradigm where two parties hold different feature spaces for the same set of overlapping entities. For example, a bank holds financial history and a retailer holds purchase history for the same customers. The training uses Entity Alignment techniques without exposing identifiers, and splits the neural network architecture across the parties to compute gradients without sharing raw features.

• Split Neural Networks: The model is physically partitioned between participants.
• Private Set Intersection: Cryptographic protocols to identify overlapping entities without revealing non-overlapping ones.