In the field of computer vision deep neural networks are proven and tested solutions for complex problems. Depth reconstruction from stereo images and estimation of optical flow from image sequences is solved most accurate by deep learning based methods. The extension of optical flow to three spatial dimensions is called scene flow. Deep neural networks for scene flow estimation outperform classical energy functional minimisation methods. The application of neural network solutions for scene flow estimation in security-critical applications such as autonomous driving or robot-assisted surgery calls for an in-depth evaluation of these systems. The manipulation of network outputs with adversarial attacks was first uncovered for object classification networks. Adversarial attacks aim at introducing imperceptible perturbations to input images to cause erroneous network outputs. Recent research could reveal the low adversarial robustness of state-of-the-art stereo matching and optical flow neural network solutions. In this thesis a framework to generate a targeted constrained global adversarial attack on scene flow neural networks (GSFA) is developed. Multiple different attack types which add perturbations to specific types of inputs or at different stages of the networks processing pipeline are introduced. The attack types are applied to the state-of-the-art scene flow estimation network RAFT-3D. The effects of GSFA regarding scene flow estimation accuracy and perturbation size of inputs is analysed using RAFT-3D and two scene flow benchmark datasets. The results of various experiments proof that RAFT-3D shows the same vulnerabilities to adversarial attacks as optical flow and stereo matching networks. Constraints on perturbation sizes effectively keep perturbations imperceptible or hardly perceptible, while scene flow estimations approach a defined zero scene flow target.