Editing Federated Learning (section)

== <span style="color: #FFFFFF;">Understanding</span> ==
The intuition behind federated learning: instead of asking a thousand hospitals to share patient records (legally and ethically fraught), you send a copy of the model to each hospital, each hospital trains it on their local patients, and each sends back only the model's updated weights. The server averages all these updates to produce a better global model. No patient record ever left any hospital.

'''The FedAvg algorithm''':
1. Server initializes global model weights w_0
2. Server sends w_t to a subset of K clients
3. Each client k trains on local data for E epochs → produces w_{t+1}^k
4. Server aggregates: w''{t+1} = Σ (n''k/N) · w_{t+1}^k (weighted average by dataset size)
5. Repeat for T rounds

'''The non-IID problem''' is FL's central challenge. If a smartphone's photos are taken mostly at night, its local model update will be biased toward night photography. Hospital A serves elderly patients; Hospital B serves pediatric patients. Their local models will diverge, and naive averaging may harm performance on each individual distribution.

'''Communication bottleneck''': Each round requires clients to transmit model weights — potentially hundreds of MB for large models. With limited bandwidth (mobile devices, rural hospitals), this is a critical constraint. Solutions include gradient sparsification (transmit only the largest gradients), quantization (reduce precision), and local distillation (compress knowledge before transmission).

'''Privacy guarantees''': Even without sharing raw data, shared gradients can reveal information. Gradient inversion attacks can reconstruct training images from gradients. Differential privacy adds calibrated noise to updates: the server never learns more than (ε, δ)-DP bounds about any individual's data.
</div>

<div style="background-color: #8B0000; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">