Adversarial Machine Learning: A Deep Dive

Author(s): Rohan Rao

Originally published on Towards AI.

A Cyber attack — Manipulating machine learning models

This member-only story is on us. Upgrade to access all of Medium.

Image created by author in canva

Today morning, I suddenly had a thought that if we are using Machine Learning models at such a huge scale, how are the vulnerabilities checked in the models itself?

Little bit searching and I found that there is something called as Adversarial Machine Learning — sounds exciting!

Image credits: https://www.researchgate.net/figure/An-adversarial-machine-learning-The-upper-layer-represent-the-traditional-machine_fig1_365747650

Adversarial Machine Learning or simply AML is a subfield of Artificial Intelligence that explores how to manipulate a machine learning model.

It can also be termed as “Cyber-attack” to fool a model with unwanted inputs.

Look at the figure for an example:

Image credits: https://www.researchgate.net/figure/An-illustration-of-machine-learning-adversarial-examples-Studies-have-shown-that-by_fig1_324055823

Studies shows that by adding a very tiny change to the data can lead the machine learning model make incorrect decisions, resulting wrong prediction.

Another example:

Image credits: https://www.educba.com/adversarial-machine-learning/

This seems concerning, isn’t it? Let’s see its key concepts first to have better understanding.

There are many types of Adversarial Attacks :

This attacks usually happen when the attacker can control the entire model’s architecture. He has controls to training data, weights, and parameters.

This happens when the attacker has limited knowledge of the model. He cannot access the model’s internal architecture. He can only query the training data, parameters etc and… Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI