Testable Sanitization via Boundary GeometryResearch Paper
A classifier trained or supplied by an untrusted party may contain a backdoor: a hidden rule that changes its label on inputs perturbed in a way only the adversary knows. Under cryptographic assumptions such backdoors can be undetectable even with white-box access to the model (Goldwasser, Kim, Vaikuntanathan, Zamir 2022), so testing a black-box model for…