Houdini: Fooling Deep Structured Prediction Models
This addresses the robustness evaluation of learning machines for tasks with combinatorial and non-decomposable measures, though it is incremental as it extends adversarial methods to structured prediction.
The paper tackled the problem of generating adversarial examples for deep structured prediction models by introducing Houdini, a flexible approach tailored to the final performance measure, resulting in higher success rates and less perceptible perturbations across applications like speech recognition and semantic segmentation.
Generating adversarial examples is a critical step for evaluating and improving the robustness of learning machines. So far, most existing methods only work for classification and are not designed to alter the true performance measure of the problem at hand. We introduce a novel flexible approach named Houdini for generating adversarial examples specifically tailored for the final performance measure of the task considered, be it combinatorial and non-decomposable. We successfully apply Houdini to a range of applications such as speech recognition, pose estimation and semantic segmentation. In all cases, the attacks based on Houdini achieve higher success rate than those based on the traditional surrogates used to train the models while using a less perceptible adversarial perturbation.