Neural networks have achieved peak performance in many computer vision tasks over the past decade, with image classification and pattern recognition among the first tasks to be mastered. Subsequent papers showed that the robustness of these networks is not yet fully understood. Nowadays, the task of estimating the optical flow is similarly best performed by flow networks. Attacks to distorted flow fields employ, for example, so-called adversarial patches customized to the various neural networks. The impact of the patch attacks depended on the underlying architecture of the network but was broadly successful. Classical methods seem to be immune to these attacks. However, one type of attack is insufficient to conclude robustness. So far, no study has tailored an attack on classical methods, but this thesis is the basis for an attack on the Horn and Schunck method. We present a method to derive the estimated optical flow with respect to the input images, which is the gradient for the optical flow. The beginning is the variation approach of Horn and Schunck, which sets up an energy functional. This functional’s minimizer must satisfy the Euler-Lagrange equations, which are set up as a system of linear equations. The solution to the linear system is the optical flow estimate. The implicit function theorem expresses that there must be a function rule from the input images to the resulting flow field in the neighbourhood of the solution. Furthermore, it provides a formula for the derivative function of this mapping. With this formula, the Horn and Schunck method can be derived with respect to the input images. This thesis aims to help understand the discretization of the Euler-Lagrange equations, the composition of the resulting linear equation system and the usage of the implicit function theorem. The work can later be used to generate images that distort the flow field estimated by the Horn and Schunck method. We expect that the classical method is not robust to such a tailored attack.