- Abstract: Despite their enormous success, there is still no solid understanding of deep neural network’s working mechanism. As such, researchers have demonstrated DNNs are vulnerable to small input perturbation, i.e., adversarial attacks. This work proposes the effective path as a new approach to exploring DNNs' internal organization. The effective path is an ensemble of synapses and neurons, which is reconstructed from a trained DNN using our activation-based backward algorithm. The per-image effective path can be aggregated to the class-level effective path, through which we observe that adversarial images activate effective path different from normal images. We propose an effective path similarity-based method to detect adversarial images and demonstrate its high accuracy and broad applicability.