Towards Verifiable Specifications for Neural Networks in Autonomous Driving

Autonomous driving functions increasingly incorporate machine learning methods for perception, environment modeling and decision making. The ever-present danger of unintended behavior of neural networks has given cause for concern in safety-critical applications and has spurred research towards the establishment of certified properties. In our work we present a general approach for exploring guaranteed properties of neural networks through a specific use case. We first methodically specified several desired properties for an ideal traffic sign classifier. Then we applied a mixed integer linear programming verification method to certify (or conversely, falsify) that our pre-trained rectified linear unit (ReLU) activation network will satisfy given properties under all conditions. We established through relaxations of our initial specification that our model does in fact happen to guarantee certain (but not all) properties at a different threshold. We thus showed that it is possible to establish the degree to which non-trivial and humanly meaningful properties are inherently guaranteed by an arbitrary pixel space ReLU network. We believe this approach can be a helpful addition in comprehensively evaluating the reliability of a network via quantifying its degree of compliance to an ideal specification.