A very quick introduction to Adversarial Perturbation in Images

Making sure that image classifier are robust is absolutely crucial to improving performance

2 min readOct 31, 2023

If you’re interested in Computer Vision, it helps to understand Adversarial Perturbation.

At its core, it’s a technique where slight modifications are added to the input data, leading our models to make incorrect predictions. The goal is to modify input images in ways that are imperceptible to humans but completely break image classifiers.

How is it implemented? Through a process of optimization, adversaries find these perturbations by maximizing the model’s error on the modified input.

An intriguing fact: Most adversarial attacks exploit the inductive biases of ConvNets. These biases, inherent to the model’s design, can be taken advantage of, resulting in misclassification. These work by attacking the ‘fragile features’ that Deep Learners would extract, throwing off the entire classifer.

However, while ConvNets might have their vulnerabilities, it’s essential to note that the newer Vision Transformers (ViTs) aren’t immune either. They’re susceptible to their own unique set of attacks. This area has not been explored in as much detail because CNNs have been the focus in Vision Research, but with the rise of multi-modal models based on Transformers- it is important to understand them.

The image below refers to the one-pixel attack, where changing a single image completely broke SOTA classifiers.

For more details, sign up for my free AI Newsletter, AI Made Simple. AI Made Simple- https://artificialintelligencemadesimple.substack.com/

If you liked this article and wish to share it, please refer to the following guidelines.

If you find my writing useful and would like to support my writing- please consider becoming a premium member of my cult by subscribing below. Subscribing gives you access to a lot more content and enables me to continue writing. This will cost you 400 INR (5 USD) monthly or 4000 INR (50 USD) per year and comes with a 60-day, complete refund policy. Understand the newest developments and develop your understanding of the most important ideas, all for the price of a cup of coffee.

Support AI Made Simple