Federated Learning vs. De-Anonymization -- Protecting Your Data

Sun, 16 Jun 2024 14:15:39 +0530

Federated Learning vs. De-Anonymization: Protecting Your Data

In simple terms, federated learning is a technique where a machine learning model is shared with different nodes, clients, or data collectors (like IoT devices). Each client device trains the model locally using its own data and then sends the updated parameters and weights to a central model. This central model then aggregates these updates, learns from them, and shares the new parameters back with the client models.

Here are some terms used in this process:

Client: Total number of devices involved in the federated learning process.
Active Client: Some clients are selected to perform training based on certain conditions such as stable network connection, charging status, etc.
Round: The number of times each active node will perform training locally and then share the final parameters with the central model.

Even when data is made anonymous, personal information can still be revealed. Two cases illustrate this issue:

Case 1: Revealing the Identity of Netflix Users

Netflix shared anonymous viewing data for a competition to enhance its recommendation system. Yet, researchers managed to uncover the identities of individuals by linking this data with publicly available IMDb ratings. By comparing the viewing habits of a small number of users in both datasets, the researchers were able to identify specific people and their viewing preferences. This raised serious concerns about the reliability of anonymization methods.

Case 2: Strava Heat Map Exposing Military Locations

Strava, a fitness-tracking app, creates a map showing the activity levels of its users around the world. Although the data is anonymous, it has accidentally revealed sensitive military locations. Researchers discovered that by studying the map, they could figure out how people move in remote areas. In some cases, this helped them locate military bases and track the movements of personnel, disclosing important information.

Measures to Prevent Information Leaks in Federated Learning

To mitigate the risks associated with data de-anonymization, several techniques can be employed:

Stochastic Gradient Descent (SGD): SGD is a technique used in machine learning to improve a model by adjusting weights based on the gradient of the loss function. By adding noise to the gradient calculation, it can enhance privacy protection by making it more difficult to extract exact data points.
Differential Privacy: Differential privacy means sharing information about a group without revealing information about individuals. It ensures that adding or removing a single data point doesn’t significantly change the results, protecting individual privacy.
Federated Averaging: Federated averaging is a method used in federated learning where multiple devices work together to train a model without sharing their raw data. Each device calculates updates to the model based on its local data. Only the updates, not the data itself, are shared and averaged to improve the global model. This approach helps maintain data privacy across decentralized systems.

Conclusion

Protecting personal information in the age of big data is a tough job that requires advanced techniques. The example of Netflix and Strava shows that even when data is anonymized, it can still be at risk of de-anonymization attacks. By using methods like stochastic gradient descent, differential privacy, and federated averaging, organizations can improve their data protection measures and better protect individual privacy.

References

How To Break Anonymity of the Netflix Prize Dataset Paper
Anonymity and the Netflix Dataset - An Essay
Heat Marks the Spot: De-Anonymizing Users’ Geographical Data on the Strava Heatmap - Paper
Lecture 23 — Differentially Private Machine Learning - By Prof. Gautam Kamath - Lecture Notes

Introduction to Federated Learning

Sat, 16 Mar 2024 14:15:39 +0530

Intro to FL(Federated Learning)

The Problem with Current Machine Learning

Data goes to your model.

Getting data from different sources can change the true nature and authenticity of the data. It also creates worries about sharing and keeping the collected data private.
Today’s machine learning algorithms have data stored in one central place. This means the authenticity and originality of the data can be altered, and the data can be fake or synthetic.

The Solution: Federated Learning (FL)

In FL, your model goes to the data.

In simpler terms, your model is shared and spread across the different data sources, and your model is trained on the data on individual devices or machines. All the updated model weights are then combined into a central model.
There is a main central model that communicates with other models distributed across many client devices. The client model is trained on the client’s system multiple times in rounds so that it can handle different types of data. After addressing data issues, the distributed models share their final weights with the central model.

For example, people from different regions use Google’s keyboard, and they get suggestions in their local languages.

Another example we’re all familiar with is next-word predictions on our phone keyboards. The suggestions are based on our typing habits.

Another use case is detecting anomalies in industrial IoT devices, as different IoT devices work differently based on their surroundings and characteristics, needing different maintenance predictions from other IoT devices.

How FL is Effective

There are various methods to use data from different sources and train models, but Federated Learning is a proven and widely used technique. It solves privacy issues and the risk of exposing personal data by not taking data from clients.

For data generalization, having many different forms of data can reduce the performance of a regular machine-learning model. To avoid this, our model should be able to handle all kinds of data.
The machine learning model should use less computing power so it doesn’t overload client devices and machines.
The model should be trained in multiple rounds to train efficiently by handling the different types of data and resource availability on clients’ devices.

source

Reading Material link

Research | Ukant Jadia