Introduction to Adversarial Attack In Computer Vision

Author(s): Vincent Liu

Originally published on Towards AI.

Source: image by author. Video source: DAVIS¹

Since we started to leverage the power of models in data science, the digital world has been evolving at an incredible speed. Nowadays we have a variety of models based on text, audio, image, and other domain-specific type of data. The community put in the effort to improve the models in terms of efficiency and accuracy.

At the MIT Spam Conference in January 2004, it was disclosed that a machine learning model could suggest a single word and put it in an email to bypass other spam mail filters. Imagine how incredible it is to know that adding a word to an email can trick the advanced mail filters at the time. The term “adversarial attack” has come under scrutiny within the community since this issue emerged.

An adversarial attack aims to mislead the model’s prediction by introducing imperceptible perturbation to the input. An example of an adversarial attack on segmentation is shown in the image at the top. The first row displays the image and corresponding predicted mask; the second row is the perturbed result. It can be seen that the difference in the input images is negligible, while the inconsistency between the masks is… 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