Privacy-Preserving AI: Federated Learning Simulator

Federated Learning

Privacy budget ε 1.00

Client devices 6

Round speed 1.0×

Global accuracy

Rounds completed

Privacy spent (Σε)

0.0

Update noise σ

–

How it works

In federated learning, raw data never leaves a client device. Each device trains a local copy of the model on its own private data, then sends only the resulting model *update* — not the data itself — to a central server. Differential privacy adds calibrated random noise to each update before it leaves the device, so no single individual's data can be reverse-engineered from what the server sees. The server simply averages every noised update into a new global model (FedAvg) and broadcasts it back out.

noise scale:  b = Δf / ε        (smaller ε → more noise, more privacy)
FedAvg:       w_global = (1/N) Σ w_i
composition:  ε_total = ε_round × rounds

Privacy budget ε — how much noise each client adds to its update. Low ε means strong privacy but a noisier, less accurate global model; high ε trades privacy for accuracy.
Client devices — how many independent devices contribute local updates each round; more clients smooth out noise in the aggregate.
Round speed — how fast the train → noise → send → aggregate → broadcast cycle repeats.
Show raw-data risk — overlays what an insecure "send the raw data directly" pipeline would look like, for comparison with the protected federated path.

Real-world relevance: this is exactly how hospitals train shared diagnostic models without pooling patient records (HIPAA/GDPR compliance), and how keyboard apps like Gboard improve next-word prediction without ever uploading what you typed.

Global model (server) Client device Noised update (safe) Raw data (insecure demo)

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.