As self-driving cars develop, safety measures that ensure resilience against adversarial attacks are necessary. Adversarial attacks can be combatted by using heterogeneous convolutional neural networks (CNNs) and a majority voting system for the outputs of CNNs. Our program explores the design space about the voting system. We devise two voting schemes: Voter1, a majority voting program, and Voter2, an OR voting program. The simulation program uses two state-of-the-art CNNs and determines the desired precision of the third neural network for the voting system to have precision and recall of at least 75%. Our evaluation shows that the precision of the overall system using Voter1 is higher than 75% when the precision of the 3rd network reaches ~90%. Voter1 leads to a higher accuracy than Voter2. Thus, Voter1 can further be applied to humanoid robots and critical systems to enhance resilience in fields such as healthcare, where high precision is critical.