The paper proves that satisfiability of binarized neural networks is NP-complete by reducing it to SAT. It also shows that uniform image occlusion leads to a piecewise-constant output structure, allowing polynomial-time robustness verification.