fedcourse24_lab4

Fourth Lab Report — Federated Learning & Data Privacy

Université Côte d’Azur – MSc Data Science & Artificial Intelligence
Author: Zeev Weizmann
November 2025

---

Overview

This project investigates label‑flipping attacks in Federated Learning and evaluates the robustness of FedAvg and coordinate‑wise median aggregation under both IID and non‑IID data distributions.
Experiments include malicious client simulation, robust aggregation, and performance comparison.

---

Visual Results

IID — No Defense

IID — Median Defense

Non‑IID — With and Without Defense

---

Exercise 1 — Label‑Flipping Attacks

Data Generation

rm -r mnist/all_data
python generate_data.py   --dataset mnist   --n_clients 10   --iid   --frac 0.1   --save_dir mnist   --seed 1234

Malicious Learner

A custom learner flips labels by +1 (mod 10):

y = (y + 1) % 10

Running Experiments

for p in 0 0.1 0.3 0.5; do
python train.py   --experiment mnist   --n_rounds 25   --prop $p   --n_clients 10 ...
done

Findings

• 0–10% malicious clients: model remains stable.
• 30%: noticeable degradation.
• 50%: the global model collapses.

FedAvg is highly vulnerable to label‑flipping attacks.

---

Exercise 2 — Median Defense

Median Aggregator

Coordinate‑wise median is used:

median_state[k] = torch.median(stacked, dim=0).values

Findings

• Works well up to 30% malicious clients.
• Fails when attackers reach 50% because the median gets dominated by adversarial updates.

Median is effective only when honest clients form a coherent majority.

---

Exercise 3 — Non‑IID Case

Findings

Median becomes worse than FedAvg under non‑IID:

• Honest clients produce heterogeneous updates.
• Median treats legitimate differences as noise.
• The global model under‑trains and accuracy drops quickly.

FedAvg tolerates gradient heterogeneity better and therefore performs higher in the non‑IID setting.

---

Links

Repository SSH clone: git@github.com:ZeevWeizmann/fedcourse24_lab4.git
Repository HTTPS: https://github.com/ZeevWeizmann/fedcourse24_lab4
Project website: https://zeevweizmann.github.io/fedcourse24_lab4/
PDF report: https://github.com/ZeevWeizmann/fedcourse24_lab4/blob/7921d96790bb66a5dba7b8a1bb147a7b4f46842c/Lab_Report%20(1).pdf