Adversarial example attack endangers the mobile edge systems such as vehicles and drones that adopt deep neural networks for visual sensing. This paper presents {\em Sardino}, an active and dynamic defense approach that renews the inference ensemble at run time to develop security against the adaptive adversary who tries to exfiltrate the ensemble and construct the corresponding effective adversarial examples. By applying consistency check and data fusion on the ensemble's predictions, Sardino can detect and thwart adversarial inputs. Compared with the training-based ensemble renewal, we use HyperNet to achieve {\em one million times} acceleration and per-frame ensemble renewal that presents the highest level of difficulty to the prerequisite exfiltration attacks. Moreover, the robustness of the renewed ensembles against adversarial examples is enhanced with adversarial learning for the HyperNet. We design a run-time planner that maximizes the ensemble size in favor of security while maintaining the processing frame rate. Beyond adversarial examples, Sardino can also address the issue of out-of-distribution inputs effectively. This paper presents extensive evaluation of Sardino's performance in counteracting adversarial examples and applies it to build a real-time car-borne traffic sign recognition system. Live on-road tests show the built system's effectiveness in maintaining frame rate and detecting out-of-distribution inputs due to the false positives of a preceding YOLO-based traffic sign detector.