Deep Neural Networks for optical flow estimation achieve exceeding performances on published datasets. However successful and practically applicable patch attacks in the past and the lack of robustness guarantees for published networks demonstrate the importance to further study their resilience against manipulated inputs. The recent desire to deploy optical flow networks in security-critical applications like autonomous driving or robot navigation further amplifies this problem. Recently, the suitability of localized pre-processing defenses against adversarial patch attacks for optical flow networks was examined and a specialized defense was proposed. However an extensive robustness assessment of the defended systems using adaptive attacks was missing. Furthermore results about adaptive attacks on defended networks are currently rather restricted to classification networks. In this thesis we devise adaptive adversarial patch attacks against the optical flow network FlowNetC when it is defended by the specialized defenses Inpainting with Laplacian Prior~(ILP) and Local Gradients Smoothing~(LGS). We provide empirical evidence that our adaptive white-box attacks increase the efficiency of injected patches significantly compared to the attacks considered in their initial evaluation. Our attacks introduce serious distortions in the flow field estimation of defended networks. Additional contributions are the implementation of a flexible training pipeline and the reimplementation of the Inpainting with Laplacian Prior defense according to its description in the original publication.