Editing Federated Learning (section)

== <span style="color: #FFFFFF;">Analyzing</span> ==
{| class="wikitable"
|+ Cross-device vs Cross-silo FL
! Aspect !! Cross-device (Smartphones) !! Cross-silo (Hospitals/Banks)
|-
| Number of clients || Millions || Tens to hundreds
|-
| Data per client || Small, highly non-IID || Large, moderately non-IID
|-
| Participation || Intermittent, unreliable || Reliable, scheduled
|-
| Communication || Limited bandwidth (mobile) || High bandwidth (datacenter)
|-
| Trust model || Untrusted clients || Semi-trusted institutions
|-
| Deployment || Google Keyboard, Apple Siri || Medical research consortia, banking
|}

'''Key challenges and failure modes:'''
* '''Non-IID degradation''' — With highly heterogeneous client data, FedAvg can converge to a poor global model or oscillate. FedProx (adds proximal term to keep local models close to global) and SCAFFOLD (variance reduction) address this.
* '''Client stragglers''' — If the server waits for slow clients, training is bottlenecked. Asynchronous FL or ignoring slow clients introduces gradient staleness.
* '''Model poisoning and backdoor attacks''' — Malicious clients inject backdoors (e.g., "if input contains a specific trigger, classify as target class") that survive averaging. Defense: Krum, median/trimmed-mean aggregation, anomaly detection on updates.
* '''Gradient inversion''' — Large batch gradients can be used to reconstruct training data. Mitigation: secure aggregation (cryptographic) + differential privacy (statistical).
* '''Concept drift''' — Client data distributions change over time; the global model becomes stale for some clients. Personalization and continual learning help.
</div>

<div style="background-color: #483D8B; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">