In this talk, I will present 1) a framework to systematically analyze convolutional neural networks and 2) a counterexample-guided data augmentation scheme. The analysis procedure comprises a falsification framework in the system-level context. The falsifier is based on an image generator that produces synthetic pictures by sampling in a lower dimension image modification subspace. The generated images are used to test the CNN and expose its vulnerabilities. The misclassified pictures (counterexamples) are then used to augment the original training set and hence improve the accuracy of the considered model. The talk focuses on case studies of object detection in autonomous driving based on deep learning.