SecureCROWN: Privacy-Preserving Robustness Verification for Neural Networks

The paper introduces SecureCROWN, a privacy-preserving framework for certified neural-network robustness verification based on secure two-party computation (2PC). A model owner and a data owner jointly compute certified robustness bounds while revealing only the final result under the semi-honest security model. To address data-dependent branching in Linear Bound Propagation, SecureCROWN reformulates conditional logic as continuous arithmetic operations. It also uses Newton–Raphson refinement to improve numerical stability. The authors report exact agreement with plaintext verification and runtimes of 0.1–200 seconds across model sizes and LAN/WAN settings.