Verifying Global Robustness Faster: Abstractions, Relaxations, and Incremental Analysis

Global robustness enables identifying, at inference time, the inputs that may be susceptible to adversarial attacks. However, its verification doubles the exponent of local robustness, which has kept it far less studied. Inspired by the abstractions that scaled local robustness, we propose lightweight abstractions for scaling global robustness, defined over two copies of a network. Most of them accelerate the verification without over-approximation, which would otherwise flag more inputs as susceptible to adversarial attacks. The other is a novel conditional-triangle relaxation that extends the popular triangle approximation to global robustness while reducing its over-approximation error. We also show, for the first time, how to perform incremental verification, shown to scale local robustness, for global robustness. We introduce BLEND, which scales exact global robustness verification by combining these complementary abstractions, and supports incremental verification, leveraging a previous network’s verification to speed up that of a revised one.

M.Sc. student under the supervision of Prof. Dana Drachsler Cohen.