© 2018 Strange Loop
Deep learning using neural networks has gained incredible popularity in recent years; and has quickly been introduced into software applications and business processes. Now, using open-source and commercial APIs, developers with little experience in predictive modeling can perform tasks like face and identity recognition; emotion detection; and object classification. What was previously impossible can now be accomplished with less than 10 lines of code.
This widespread dissemination of deep learning is a huge opportunity to democratize AI - but offers considerable potential for exploitation. Models can only be as good as the data that feeds them; and, in this presentation, we will explore several ways to break pre-built models using minimal disruption (ex: single-pixel attacks; random noise). We will also attempt to reverse-engineer models and tailor our inputs for specific classifications.
Paige Bailey is a senior Cloud Developer Advocate at Microsoft specializing in Machine Learning and Artificial Intelligence. Prior to joining Microsoft, Paige was employed as a data scientist and predictive modeler in the energy industry - high-consequence deep-water projects, and onshore statistical plays. Paige is on the committee for SciPy and JupyterCon, is a Python instructor on EdX, and is writing a book on machine learning at scale.