Optical flow estimation is used in many applications of modern life. While traditional methods rely on knowledge-based energy minimization, state-of-the-art techniques use deep neural nets to approximate the flow. Neural nets are shown to be vulnerable to perturbed input images. Recent research detected these vulnerabilities for accurate flow estimation methods, such as RAFT [66]. Contrary, traditional methods have not been investigated using such adversarial samples yet. In this bachelor thesis, we propose a novel adversarial attack tailored to the method of Horn and Schunck [27]. Instead of minimizing the originally defined energy to obtain the optical flow, the target flow is substituted into the energy. To obtain perturbed images, we consider the initial images as unknowns and use an optimization approach based on automatic differentiation to solve for perturbations. When applied to the perturbed images, the Horn-Schunck method returns values close to the target flow. As a second key aspect of the thesis, we transfer the energy minimization of Horn and Schunck to the loss function of RAFT for a neural net that resembles the traditional method more closely. Our attacks on the method of Horn and Schunck reveal the method to be more robust than RAFT. When the produced perturbations cause meaningful change in the resulting flow, they are systematic and visible to the human eye. For neural nets, these perturbations are often unrecognizable. The modifications to the loss function of RAFT lead to a large drop in performance and a significant increase in robustness to adversarial attacks. A direct combination of data- and knowledge-driven losses results in a drop in accuracy and robustness.