Stealing Models from the Cloud

Decision Boundaries

Understanding model decision boundaries is crucial for efficiently recovering machine learning models through iterative queries. Florian Tramèr explains that many machine learning services provide probabilistic outputs, which can be exploited to recover model parameters by solving equations, akin to training a model 1. This process involves querying random points and exploring regions between differently classified points to approximate decision boundaries.

By finding points closer to the decision boundary, you can learn where the model makes its decisions.

--- Florian Tramèr

This method allows for a better understanding of how models split classes, even in high-dimensional spaces 2.

   

Function Boundaries

Learning function boundaries involves understanding how models function by identifying decision boundaries, which are mathematically equivalent to the extraction process. Florian Tramèr notes that unlike traditional machine learning, where the function doesn't exist per se, extraction targets a well-defined model with clear decision boundaries 3. This makes learning easier as there is no noise, and class probabilities aid in recovering the model.

Recovering a model given access to these exact probability outputs turns out to be much, much easier.

--- Florian Tramèr

These insights highlight the efficiency of learning decision boundaries in model extraction 3.