On Your Ocean's 11 Team, I'm the AI Guy (or Girl)

The talk uses a casino heist as a frame for looking at AI security. The speaker explains why facial recognition is the most important AI system in a casino, since surveillance is built into how the venue operates and card counting itself is not illegal. Interviews with organisations suggest that many firms run AI systems but few can describe how they secure them.

The middle of the talk covers adversarial machine learning, where small crafted changes to an input cause a model to misclassify it. Open-source face recognition models tend to learn similar internal structure, so a surrogate model can stand in for a target. The method places adversarial regions around a subject rather than altering the subject itself. A code demo on an open-source model showed match confidence dropping by roughly 40 percent, and results varied widely between models because machine learning outputs are probabilistic.

The speaker argues that AI security should be discussed in terms of risk and the highest-value assets, not as a simple count of attacks and defences. Disruption-style attacks that reliably break a model may matter more to an operator than a precise deception. Ownership of AI security is often unclear, sitting somewhere between IT, data science and cyber teams, and regulation rarely requires robustness testing.