Federated Learning vs Centralized Learning in Cloud Computing

🔷 Introduction

Machine learning models for cloud resource management are traditionally trained in a centralized manner, where all data is aggregated into a single server. However, with growing concerns over privacy, scalability, and data ownership, Federated Learning (FL) has emerged as a viable alternative.

This article presents a detailed comparison of Centralized Learning vs Federated Learning in cloud environments.

🔷 Centralized Learning Overview

In centralized learning:

All workload data (CPU, memory, network) is sent to a central server
Model is trained on aggregated data
Results are distributed back

✔ Advantages

High accuracy (global data visibility)
Simpler implementation

❌ Limitations

Data privacy risks
High communication overhead
Single point of failure

🔷 Federated Learning Overview

In federated learning:

Data remains on local nodes
Models are trained locally
Only model updates are shared

✔ Advantages

Privacy-preserving
Scalable
Distributed intelligence

❌ Limitations

Slightly lower accuracy
Communication rounds required
Heterogeneous client challenges

🔷 Architectural Comparison

Feature	Centralized Learning	Federated Learning
Data Location	Central server	Distributed
Privacy	Low	High
Scalability	Limited	High
Communication	High (raw data)	Moderate (model updates)
Fault Tolerance	Low	High

🔷 Mathematical Perspective

Centralized optimization: $\min_w F(w)$ wminF(w)

Federated optimization: $\min_w \sum_{k=1}^{K} p_k F_k(w)$ wmink=1∑KpkFk(w)

🔷 Experimental Insights (Align with your work)

Centralized models:
- Lower MAE, RMSE
- Better convergence
Federated models:
- Slight accuracy trade-off
- Better scalability and privacy
Trust-weighted FL:
- Handles malicious clients
- Improves robustness

🔷 Use Cases in Cloud Computing

Centralized Learning:

Small-scale cloud systems
Controlled environments