Abstract-Deep neural networks have achieved near-human accuracy levels in various types of classification and prediction tasks including images, text, speech, and video data. However, the networks continue to be treated mostly as black-box function approximators, mapping a given input to a classification output. The next step in this human-machine evolutionary processincorporating these networks into mission critical processes such as medical diagnosis, planning and control -requires a level of trust association with the machine output.Typically, statistical metrics are used to quantify the uncertainty of an output. However, the notion of trust also depends on the visibility that a human has into the working of the machine. In other words, the neural network should provide humanunderstandable justifications for its output leading to insights about the inner workings. We call such models as interpretable deep networks.Interpretability is not a monolithic notion. In fact, the subjectivity of an interpretation, due to different levels of human understanding, implies that there must be a multitude of dimensions that together constitute interpretability. In addition, the interpretation itself can be provided either in terms of the lowlevel network parameters, or in terms of input features used by the model. In this paper, we outline some of the dimensions that are useful for model interpretability, and categorize prior work along those dimensions. In the process, we perform a gap analysis of what needs to be done to improve model interpretability.

Differential privacy (DP) is a widely accepted mathematical framework for protecting data privacy. Simply stated, it guarantees that the distribution of query results changes only slightly due to the modification of any one tuple in the database. This allows protection, even against powerful adversaries, who know the entire database except one tuple. For providing this guarantee, differential privacy mechanisms assume independence of tuples in the database-a vulnerable assumption that can lead to degradation in expected privacy levels especially when applied to real-world datasets that manifest natural dependence owing to various social, behavioral, and genetic relationships between users. In this paper, we make several contributions that not only demonstrate the feasibility of exploiting the above vulnerability but also provide steps towards mitigating it. First, we present an inference attack, using real datasets, where an adversary leverages the probabilistic dependence between tuples to extract users' sensitive information from differentially private query results (violating the DP guarantees). Second, we introduce the notion of dependent differential privacy (DDP) that accounts for the dependence that exists between tuples and propose a dependent perturbation mechanism (DPM) to achieve the privacy guarantees in DDP. Finally, using a combination of theoretical analysis and extensive experiments involving different classes of queries (e.g., machine learning queries, graph queries) issued over multiple large-scale real-world datasets, we show that our DPM consistently outperforms state-of-the-art approaches in managing the privacy-utility tradeoffs for dependent data. Permission to freely reproduce all or part of this paper for noncommercial purposes is granted provided that copies bear this notice and the full citation on the first page. Reproduction for commercial purposes is strictly prohibited without the prior written consent of the Internet Society, the first-named author (for reproduction of an entire paper only), and the author's employer if the paper was prepared within the scope of employment.

No abstract

Saliency maps are a popular approach to creating post-hoc explanations of image classifier outputs. These methods produce estimates of the relevance of each pixel to the classification output score, which can be displayed as a saliency map that highlights important pixels. Despite a proliferation of such methods, little effort has been made to quantify how good these saliency maps are at capturing the true relevance of the pixels to the classifier output (i.e. their “fidelity”). We therefore investigate existing metrics for evaluating the fidelity of saliency methods (i.e. saliency metrics). We find that there is little consistency in the literature in how such metrics are calculated, and show that such inconsistencies can have a significant effect on the measured fidelity. Further, we apply measures of reliability developed in the psychometric testing literature to assess the consistency of saliency metrics when applied to individual saliency maps. Our results show that saliency metrics can be statistically unreliable and inconsistent, indicating that comparative rankings between saliency methods generated using such metrics can be untrustworthy.

Synthesis of various nanostructured semiconductor materials and processing them for different device fabrications has been at the forefront of research for the last two decades. In comparison to spherical nanoparticles, anisotropic materials e.g. nanorods, nanowires, and nanodisks have been widely explored to obtain a better performance of the devices. In addition, it is also well-known that nanomaterials, on doping with suitable impurities, can enhance the device sensitivity and speed. Combining both, we report here the synthesis of micrometer long In2S3 nanosheets and on doping them with Cu(I), we have studied here their photoresponse properties. These nanosheets are synthesized in a high temperature colloidal method following a catalytic thermal decomposition of a single source precursor of In and S. From various TEM, HRTEM, and HAADF images the growth pattern of these sheets is investigated, and the obtained moiré fringes at the overlapped region are discussed. Finally, the comparative study of the device performance has been carried out with introducing different amounts of copper in these nanosheets.