Biblio

Adversarial examples are known as carefully perturbed images fooling image classifiers. We propose a geometric framework to generate adversarial examples in one of the most challenging black-box settings where the adversary can only generate a small number of queries, each of them returning the top-1 label of the classifier. Our framework is based on the observation that the decision boundary of deep networks usually has a small mean curvature in the vicinity of data samples. We propose an effective iterative algorithm to generate query-efficient black-box perturbations with small p norms which is confirmed via experimental evaluations on state-of-the-art natural image classifiers. Moreover, for p=2, we theoretically show that our algorithm actually converges to the minimal perturbation when the curvature of the decision boundary is bounded. We also obtain the optimal distribution of the queries over the iterations of the algorithm. Finally, experimental results confirm that our principled black-box attack algorithm performs better than state-of-the-art algorithms as it generates smaller perturbations with a reduced number of queries.

Corpora used to learn open-domain Question-Answering (QA) models are typically collected from a wide variety of topics or domains. Since QA requires understanding natural language, open-domain QA models generally need very large training corpora. A simple way to alleviate data demand is to restrict the domain covered by the QA model, leading thus to domain-specific QA models. While learning improved QA models for a specific domain is still challenging due to the lack of sufficient training data in the topic of interest, additional training data can be obtained from related topic domains. Thus, instead of learning a single open-domain QA model, we investigate domain adaptation approaches in order to create multiple improved domain-specific QA models. We demonstrate that this can be achieved by stratifying the source dataset, without the need of searching for complementary data unlike many other domain adaptation approaches. We propose a deep architecture that jointly exploits convolutional and recurrent networks for learning domain-specific features while transferring domain-shared features. That is, we use transferable features to enable model adaptation from multiple source domains. We consider different transference approaches designed to learn span-level and sentence-level QA models. We found that domain-adaptation greatly improves sentence-level QA performance, and span-level QA benefits from sentence information. Finally, we also show that a simple clustering algorithm may be employed when the topic domains are unknown and the resulting loss in accuracy is negligible.