Tag
The paper introduces a 'wrong-physics backdoor' poisoning attack on neural PDE operators, where triggered inputs cause models to output valid solutions for alternate physical parameters, highlighting a critical security vulnerability in reusable solver archives.