Blogs on robustness » Lamarr Institute

00 Blog Adilova Pflanzen Strasse - Lamarr Institute for Machine Learning (ML) and Artificial Intelligence (AI)

Plants don’t walk on the street: Using human knowledge for reliable semantic segmentation

17. November 2021

Data-driven sensor interpretation in autonomous driving can sometimes lead to very implausible predictions. A partly human-designed, partly learned rule set can make existing deep neural networks that process low-level sensor information reliable....

00 Blog weickhmann - Lamarr Institute for Machine Learning (ML) and Artificial Intelligence (AI)

Application

Priority for safe Machine Learning: On the robustness of neural networks

14. July 2021

Hostile attacks on Machine Learning prediction models, for example in autonomous driving, could lead to great (personal) damage. Defense mechanisms can be used to protect ML models and increase their robustness against attacks. ...

00 Blog Adilova 2 - Lamarr Institute for Machine Learning (ML) and Artificial Intelligence (AI)

Basics

Defense game: What can we do against adversarial examples?

12. July 2021

Machine Learning software is vulnerable to certain attacks: artificially altered input data can lead to failure of the underlying model. We present techniques for developing adversarial robustness....

00 Blog Georgiev Franken - Lamarr Institute for Machine Learning (ML) and Artificial Intelligence (AI)

Research

Robust or vulnerable – gaining more insights into neural networks with heat

01. June 2021

How easily AI algorithms can be attacked depends on the shape of the boundaries with which the neural network separates decisions from each other. This can be determined by heating up the decision boundaries....