Toward Better Image Classification
As proficient as image classifiers have become, they still fail to recognize images in the way that we might hope. They can detect airplanes accurately, but the sky plays an important role in that detection. They can detect cars, but are likely to use the road as an indication. We confirm this experimentally through the use of the first localized, imperceptible adversarial attack -- by changing pixels in the background of an image, we are able to fool image classifiers. We further contribute to adversarial research with our new adversarial training technique. By training classifiers on both normal images and images with adversarial backgrounds, we seek to maintain the state-of-the-art benchmarks for classification accuracy while creating more robust models that focus on the object itself.