Federated ML — Java-Based Distributed System

Privacy-Preserving Healthcare Analytics

Based on: Ren, Li & Wu (2024) — Privacy-Preserving Data Analysis Using Federated Learning DOI: 10.69987/AIMLR.2024.50104

---

Build

cd federated-learning-java
mvn package -q

This produces target/federated-learning-java.jar

---

Run — Distributed (Separate Terminals)

Terminal 1 — Start the Server

java -cp target/federated-learning-java.jar com.fl.runner.ServerRunner 9090 3 20
#                                                                        ^    ^  ^
#                                                                      port  min  max
#                                                                           clients rounds

Server waits until 3 hospital clients connect, then training begins automatically.

---

Terminal 2 — Hospital 1

java -cp target/federated-learning-java.jar com.fl.runner.ClientRunner 1 localhost 9090
#                                                                       ^     ^       ^
#                                                                  hospitalId host   port

Terminal 3 — Hospital 4

java -cp target/federated-learning-java.jar com.fl.runner.ClientRunner 4 localhost 9090

Terminal 4 — Hospital 7

java -cp target/federated-learning-java.jar com.fl.runner.ClientRunner 7 localhost 9090

Training starts as soon as all 3 clients connect. Each client independently trains and sends updates.

---

Run on Different Machines (Real Distributed)

Machine A — Server:     ServerRunner 9090 3 20
Machine B — Hospital 1: ClientRunner 1 <serverIP> 9090
Machine C — Hospital 4: ClientRunner 4 <serverIP> 9090
Machine D — Hospital 7: ClientRunner 7 <serverIP> 9090

---

What You See Per Terminal

Server Terminal:

[Server] Listening on port 9090...
[Server] Waiting for 3 clients...
[Server] ✓ Registered: Hospital-1  (total: 1)
[Server] ✓ Registered: Hospital-4  (total: 2)
[Server] ✓ Registered: Hospital-7  (total: 3)
[Server] 3 clients connected. Starting training!

══ ROUND 1/20 ══
[Server] Global model sent. Waiting for updates...
[Server] ← Update from Hospital-1  round=1  size=8,432 bytes
[Server] ← Update from Hospital-4  round=1  size=8,432 bytes
[Server] ← Update from Hospital-7  round=1  size=8,432 bytes
[Server] ✓ Round 1 — Accuracy: 74.32%  Loss: 0.5821
...
[Server] ✓ Round 15 — Accuracy: 91.84%  Loss: 0.2743
[Server] Model converged at round 16!
[Server] TRAINING COMPLETE

Client Terminal (Hospital-1):

[Hospital-1] Connected!
[Hospital-1] Registered. Welcome! Registered clients: 1
[Hospital-1] Waiting for training rounds...

[Hospital-1] ══ Round 1 ══
[Hospital-1] Training locally on 5,432 private samples (7 epochs)...
[Hospital-1] Local accuracy: 72.18%  (1823ms)
[Hospital-1] DP applied — ε_round=0.0041  ε_total=0.0041
[Hospital-1] → Sent encrypted update (8,432 bytes)
[Hospital-1] ← Server result: accuracy=74.32%  loss=0.5821
...
[Hospital-1] FEDERATED TRAINING COMPLETE
[Hospital-1] Final model accuracy (local test): 91.12%

---

Hospital IDs and Dataset Sizes

Hospital ID	Type	Records
0	Urban	~7,500 (used as server validation)
1	Urban	~6,500
2	Urban	~6,000
3	Suburban	~5,500
4	Suburban	~5,000
5	Suburban	~5,000
6	Rural	~4,500
7	Rural	~4,000
8	Rural	~3,500
9	Specialty	~2,500

---

System Architecture

┌─────────────────────────────────────────────────────────────┐
│                    NETWORK BOUNDARY                         │
│                                                             │
│  ┌──────────────────────────────────────────────────────┐  │
│  │              FL AGGREGATION SERVER                   │  │
│  │  - Listens on TCP port 9090                          │  │
│  │  - RSA-2048 key pair                                 │  │
│  │  - FedAvg aggregation + Byzantine filter             │  │
│  │  - Global model evaluation each round                │  │
│  └─────────────────────┬────────────────────────────────┘  │
│                         │ AES-256-GCM encrypted             │
│              ┌──────────┼──────────┐                        │
│              │          │          │                        │
│    ┌─────────┴──┐ ┌─────┴──────┐ ┌┴───────────┐           │
│    │ Hospital-1 │ │ Hospital-4 │ │ Hospital-7 │           │
│    │            │ │            │ │            │           │
│    │ Own data   │ │ Own data   │ │ Own data   │           │
│    │ Local train│ │ Local train│ │ Local train│           │
│    │ DP noise   │ │ DP noise   │ │ DP noise   │           │
│    │ Encrypt    │ │ Encrypt    │ │ Encrypt    │           │
│    └────────────┘ └────────────┘ └────────────┘           │
│                                                             │
│     Raw patient data NEVER crosses this boundary            │
└─────────────────────────────────────────────────────────────┘

---

Privacy Parameters (Ren et al., 2024 — Table 2)

Mechanism	Δ	σ	ε	δ
Gradient Perturbation	0.1	0.05	0.5	1e-5
Aggregation	0.02	0.01	0.3	1e-6
Evaluation	0.001	0.008	0.2	1e-7

---

What This Adds Beyond the Base Paper

Feature	Base Paper	This System
Language	Python	Java 11
Distribution	Simulation	Real TCP sockets
Cryptography	None	AES-256-GCM + RSA
Digital signatures	None	SHA256withRSA
Separate terminals	No	Yes
Cross-machine support	No	Yes
Cloud dependency	Yes	None