Edge-Based Learning for Improved Classification Under Adversarial Noise