Editing Federated Learning (section)

== 5.4.1 Differential Privacy (DP) ==

Differential Privacy (DP) is a formal framework ensuring that the model’s outputs (e.g., parameter updates) do not reveal individual records. In FL, DP often involves injecting calibrated noise into gradients or model weights on each client. This noise is designed so that the global model’s performance remains acceptable, yet attackers cannot reliably infer any single data sample’s presence in the training set.

A step-by-step timeline of DP in an FL context can be summarized as follows:
# Clients fetch the global model and compute local gradients.
# Before transmitting gradients, clients add randomized noise to mask specific data patterns.
# The central server aggregates the noisy gradients to produce a new global model.
# Clients download the updated global model for further local training.

By carefully tuning the “privacy budget” (ε and δ), DP can balance privacy against model utility.<sup>[1]</sup><sup>[4]</sup>