Robust Privacy: Inference-Stage Privacy through Certified Robustness

The episode discusses a paper titled "Robust Privacy: Inference-Stage Privacy through Certified Robustness." The hosts explain how applying certified robustness—making model predictions invariant within a neighborhood of input data—serv as a privacy guarantee. This method successfully mitigates attribute inference and model inversion attacks, offering quantifiable bounds that outperform traditional defenses.

Open the full episode

More from AI Radio