Adversarial content manipulation for analyzing and improving model robustness

Rakshith Shetty

Abstract: The recent rapid progress in machine learning systems has opened up many real-world applications --- from recommendation engines on web platforms to safety critical systems like autonomous vehicles. A model deployed in the real-world will often encounter inputs far from its training distribution. For example, a self-driving car might come across a black stop sign in the wild. To ensure safe operation, it is vital to quantify the robustness of machine learning models to such out-of-distribution data before releasing them into the real-world. However, the standard paradigm of benchmarking machine learning models with fixed size test sets drawn from the same distribution as the training data is insufficient to identify these corner cases efficiently. In principle, if we could generate all valid variations of an input and measure the model response, we could quantify and guarantee model robustness locally. Yet, doing this with real world data is not scalable. In this thesis, we propose an alternative, using generative models to create synthetic data variations at scale and test robustness of target models to these variations. We explore methods to generate semantic data variations in a controlled fashion across visual and text modalities. We build generative models capable of performing controlled manipulation of data like changing visual context, editing appearance of an object in images or changing writing style of text. Leveraging these generative models we propose tools to study robustness of computer vision systems to input variations and systematically identify failure modes. In the text domain, we deploy these generative models to improve diversity of image captioning systems and perform writing style manipulation to obfuscate private attributes of the user. Our studies quantifying model robustness explore two kinds of input manipulations, model-agnostic and model-targeted. The model-agnostic manipulations leverage human knowledge to choose the kinds of changes without considering the target model being tested. This includes automatically editing images to remove objects not directly relevant to the task and create variations in visual context. Alternatively, in the model-targeted approach the input variations performed are directly adversarially guided by the target model. For example, we adversarially manipulate the appearance of an object in the image to fool an object detector, guided by the gradients of the detector. Using these methods, we measure and improve the robustness of various computer vision systems -- specifically image classification, segmentation, object detection and visual question answering systems -- to semantic input variations.