Property-Directed Verification of Recurrent Neural Networks
This addresses verification challenges for recurrent neural networks, offering a property-directed method that is incremental in combining learning and model checking.
The paper tackles verifying recurrent neural networks by learning a deterministic finite automaton as a surrogate model using active automata learning, guided by a property to enable fast discovery of small counterexamples and generalization through pumping towards faulty flows.
This paper presents a property-directed approach to verifying recurrent neural networks (RNNs). To this end, we learn a deterministic finite automaton as a surrogate model from a given RNN using active automata learning. This model may then be analyzed using model checking as verification technique. The term property-directed reflects the idea that our procedure is guided and controlled by the given property rather than performing the two steps separately. We show that this not only allows us to discover small counterexamples fast, but also to generalize them by pumping towards faulty flows hinting at the underlying error in the RNN.