Deep Neural Networks (DNNs) have proven to be an alternative for object identification formultiple application areas. They are treated as a critical component for autonomouslyoperating systems and consequently crucial for many companies. Since DNNs do not behavein the same way as traditional deterministic systems, there are several challenges to cope withbefore being used in safety-critical applications. Both random and systematic failures must betaken care of, including permanent and transient faults, design faults in hardware andsoftware, and adversarial inputs. In this thesis, we will be constructing an architecture that isrobust and can detect misleading errors produced by a DNN to some extent. One way to copewith failures in DNNs is through architectural mitigation. By adding redundant and diversearchitectures, it can detect misclassification to a greater area. Convolutional neural networkarchitectures will be tested and trained using MATLAB and Simulink. The focus will be onfault-tolerant architectures. The method used in this thesis is experimental research. Theresults show that parallel architectures can detect misleading image classification better. Inaddition, and somewhat unexpected, combining three different networks gives worse resultsthan combining two networks.