[ZISC Lunch Seminar] Physical-world attacks on machine-learning (ML) + Using ML to help users make better security decisions

This talk will show why we should be concerned about the increasing use of machine-learning (ML) algorithms in safety- and security-critical applications; but also how machine learning can help users maintain their privacy.

 

First, I will show that state-of-the-art face-recognition algorithms are vulnerable to _physically realizable_ and _inconspicuous_ attacks, allowing attackers to evade recognition or impersonate specific people. We develop a systematic method to automatically generate such attacks, which are realized through printing a pair of eyeglass frames on a consumer photo printer.

 

Second, I will describe several uses of ML that help users make better decisions when using social networks, browsing the web, or creating

new passwords. Focusing on password creation, I will show how we harness neural networks to model the strength of text passwords---more quickly, more accurately, and using less space than any previous method. This makes it possible to give users accurate, actionable feedback when they are creating passwords---and our empirical results show that this feedback helps many users to create vastly stronger passwords.