Model Poisoning: Pre-Lab

Federated Learning - Attack and Defense Demonstration: Pre-Lab

Learning Objective:

In this lab, we will demonstrate the concept of federating learning as well as its susceptibility to attacks. We will also show a potential defense to one such type of attack.

Introduction to Federated Learning:

Federated learning is a type of distributed learning in which some model uses multiple devices or servers in order to train. It is a technique that is often used when privacy is a primary concern and or moving the training data to a centralized server isn't possible.

One way to attack a federated learning environment is to use a malicious device to introduce noisy updates to the model. This is what we will demonstrate.

Setting Up the Environment:

• We will be using a Jupyter notebook

• We will use a linear regression model. Our los metric will be mean squared error (MSE).

Understanding the Linear Regression Model:

• The model is initialized with a learning rate, weights, and bias.

• The prediction is based on the linear equation y = wx + b, where w is the weight and b is the bias.

• The model's performance is evaluated using the Mean Squared Error (MSE).

Local Training and Federated Learning:

• We will be simulating the use of multiple devices for the purposes of demonstration.

• We will then take an aggregate of all of these simulated devices to update the model.

• We will generate random data to simulate the distribution.

Data Preprocessing and Initial Federated Learning:

• We will use the diabetes dataset from sklearn.

• We will standardize the dataset and split it into separate blocks to simulate data on different devices.

• We will evaluate the global model's performance before and after federated learning.

Adversarial Attack: Introducing Noisy Data:

• We will simulate an attack by introducing a malicious device with noisy data.

• The poisoned data is created by adding noise to the features and targets of the dataset.

• We will measure the impact of the introduction of this noisy data.

Averaged Federated Learning with Trust Scores:

• We will implement a defense to this attack - trust scores.

• We will assign trust scores to the devices which will determine how much the model should weight that device's contribution.

• The devices with low trust scores will have little influence over model updates which should mitigate the effectiveness of the attack.

Required Materials:

• Python environment with necessary libraries (numpy, sklearn).

• Diabetes dataset from sklearn.

• A computer or server with multiple devices or simulated devices for federated learning.

Expected Results:

We expect that the global model's performance should get better after the introduction of federated learning; however, once we introduce a noisy device, we expect the global model's performance to decrease. Then, once we implement trust scores, we expect that the global model's performance will once again return near its baseline.