学术文献库

The most scalable approaches to certifying neural network robustness depend on computing sound linear lower and upper bounds for the network's activation functions. Current approaches are limited in that the linear bounds must be handcrafted by an expert, and can be sub-optimal, especially when the network's architecture composes operations using, for example, multiplication such as in LSTMs and the recently popular Swish activation. The dependence on an expert prevents the application of robustness certification to developments in the state-of-the-art of activation functions, and furthermore the lack of tightness guarantees may give a false sense of insecurity about a particular model. To the best of our knowledge, we are the first to consider the problem of automatically computing tight linear bounds for arbitrary n-dimensional activation functions. We propose LinSyn, the first approach that achieves tight bounds for any arbitrary activation function, while only leveraging the mathematical definition of the activation function itself. Our approach leverages an efficient heuristic approach to synthesize bounds that are tight and usually sound, and then verifies the soundness (and adjusts the bounds if necessary) using the highly optimized branch-and-bound SMT solver, dReal. Even though our approach depends on an SMT solver, we show that the runtime is reasonable in practice, and, compared with state of the art, our approach often achieves 2-5X tighter final output bounds and more than quadruple certified robustness.