Detecting Adversarial Attacks via Subset Scanning of Autoencoder Activations and Reconstruction Error