Reviews: A Convex Relaxation Barrier to Tight Robustness Verification of Neural Networks

The paper targets the problem of robustness verification of neural networks. This is a very popular and important problem. One of the prominent ways to deal with it is by formulating it as a nonlinear optimization problem and then relaxing its constraints to form a linear program relaxation. These relaxations are not guaranteed to return the optimal value, but they can be solved in polynomial time and provide bounds on the optimal solution. The main contributions of the paper are as follows: 1. Proposing a unified framework that generalizes all known layerwise LP relaxations, and showing their relationship (i.e., which relaxation is tighter).