Editing Federated Learning (section)

== <span style="color: #FFFFFF;">Evaluating</span> ==
Expert FL evaluation extends beyond model accuracy:

'''Privacy-utility trade-off''': Differential privacy adds noise that degrades model accuracy. Practitioners track the Pareto frontier between ε (privacy budget) and model utility. Lower ε = stronger privacy = more noise = worse model. The right operating point depends on the use case's sensitivity.

'''Communication efficiency metrics''': Total bytes transmitted per round × number of rounds. Compare baseline FedAvg vs. compressed (quantized, sparse) communication. Target: <1% accuracy drop for >10× communication reduction.

'''Fairness across clients''': Global accuracy can mask poor performance on underrepresented clients. Evaluate accuracy per client or per client cluster and use min-max or worst-case client fairness objectives.

'''Convergence analysis''': Track global validation loss over rounds. FL convergence is slower and noisier than centralized training — more rounds are needed to account for communication overhead and non-IID effects.

Expert practitioners in cross-silo FL conduct '''institutional audits''' — verifying that each participating institution's data governance policies are respected and that the FL protocol doesn't violate regulatory requirements (HIPAA, GDPR).
</div>

<div style="background-color: #2F4F4F; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">