Federated Averaging (FedAvg)

Each participating client performs multiple steps of stochastic gradient descent on its local data partition before communicating. This contrasts with traditional distributed training that requires gradient synchronization after every batch. By executing E local epochs, FedAvg significantly reduces communication rounds—a critical advantage in bandwidth-constrained radio access networks. The local update rule follows: w_k = w_k - η∇L(w_k; batch) where η is the learning rate and L is the local loss function.

The central server aggregates client updates using a weighted average proportional to each client's dataset size n_k. The global update rule is: w_global = Σ (n_k / n_total) * w_k. This weighting mechanism ensures that clients with more representative data exert proportionally greater influence on the global model, while preventing any single large dataset from dominating convergence. The approach naturally handles statistical heterogeneity across base stations with varying traffic patterns.

FedAvg operates in discrete synchronized rounds:

• Round 1: Server broadcasts current global model w_t to selected clients
• Round 2: Each client k initializes with w_t, trains for E local epochs on private data D_k, producing updated weights w_k
• Round 3: Clients transmit only model deltas Δw_k = w_k - w_t back to the server
• Round 4: Server computes weighted average of deltas and updates global model This structure minimizes uplink bandwidth consumption—critical for edge inference offloading scenarios.

A defining challenge for FedAvg is statistical heterogeneity where local datasets are non-IID. In telecom deployments, different base stations observe fundamentally different traffic distributions—urban macrocells see dense daytime demand while rural cells exhibit sparse, bursty patterns. This causes client drift: local models diverge toward their own local optima, and naive averaging can produce a global model that performs poorly on all distributions. The weight divergence is proportional to E * η * (gradient dissimilarity), requiring careful tuning of local epochs.

Each round, the server selects a fraction C of available clients for participation. This client selection strategy balances model convergence speed against communication overhead. FedAvg is inherently vulnerable to stragglers—slow devices that delay round completion. Production deployments implement timeout thresholds: if a client fails to report within T_max, its update is discarded for that round. This trades off some statistical accuracy for predictable latency, essential for zero-touch network provisioning workflows.

FedAvg provides a baseline privacy guarantee: raw training data never leaves the base station. Only mathematical weight updates are transmitted. However, these updates are not inherently cryptographically private—gradients can leak information through model inversion attacks. Production telecom deployments layer FedAvg with:

• Secure Aggregation to prevent the server from inspecting individual updates
• Differential Privacy via Gaussian noise injection before transmission
• Gradient clipping to bound per-example sensitivity This layered approach satisfies data sovereignty requirements for cross-jurisdiction deployments.