No abstract

We introduce natural adversarial examples -real-world, unmodified, and naturally occurring examples that cause classifier accuracy to significantly degrade. We curate 7,500 natural adversarial examples and release them in an ImageNet classifier test set that we call IMAGENET-A. This dataset serves as a new way to measure classifier robustness. Like p adversarial examples, IMAGENET-A examples successfully transfer to unseen or black-box classifiers. For example, on IMAGENET-A a DenseNet-121 obtains around 2% accuracy, an accuracy drop of approximately 90%. Recovering this accuracy is not simple because IMAGENET-A examples exploit deep flaws in current classifiers including their over-reliance on color, texture, and background cues. We observe that popular training techniques for improving robustness have little effect, but we show that some architectural changes can enhance robustness to natural adversarial examples. Future research is required to enable robust generalization to this hard ImageNet test set.

Aim To measure the impact of taxes and prices on alcohol use with particular attention to the different context of rising alcohol consumption in low‐ and middle‐income countries. Methods Systematic review: we searched MEDLINE, Embase, EconLit and LILACS, grey literature, hand‐searched five specialty journals and examined references of relevant studies. We considered all reviews that included studies that quantitatively examined the relationship between alcohol prices or taxes and alcohol use. At least two reviewers independently screened the articles and extracted the characteristics, methods and main results and assessed the quality of each included study. We identified 30 reviews. Results There was overwhelming evidence that higher alcohol prices and taxes were associated with lower total alcohol consumption and that price responsiveness varied by beverage type. Total own‐price elasticities of alcohol demand were consistently negative and substantial enough to be policy meaningful; total own‐price elasticities for beer, wine and spirits were found to be approximately −0.3, −0.6 and −0.65. Reviews generally concluded that higher taxes and prices were associated with lower heavy episodic drinking and heavy drinking, although the magnitude of these associations was generally unclear. Reviews provided no evidence that alcohol price responsiveness differed by socioeconomic status, mixed and contradictory evidence with respect to age and sex and limited evidence that price responsiveness in low‐ and middle‐income countries was approximately the same as in high‐income countries. Conclusions Taxes are effective in reducing alcohol use. Moreover, increasing the price of alcohol by increasing taxes can also be expected to increase tax revenue, because the demand for alcohol is most certainly inelastic.

No abstract

We consider Location-based Service (LBS) settings, where a LBS provider logs the requests sent by mobile device users over a period of time and later wants to publish/share these logs. Log sharing can be extremely valuable for advertising, data mining research and network management, but it poses a serious threat to the privacy of LBS users. Sender anonymity solutions prevent a malicious attacker from inferring the interests of LBS users by associating them with their service requests after gaining access to the anonymized logs. With the fast-increasing adoption of smartphones and the concern that historic user trajectories are becoming more accessible, it becomes necessary for any sender anonymity solution to protect against attackers that are trajectory-aware (i.e. have access to historic user trajectories) as well as policy-aware (i.e they know the log anonymization policy). We call such attackers TP-aware.This paper introduces a first privacy guarantee against TP-aware attackers, called TP-aware sender k-anonymity. It turns out that there are many possible TP-aware anonymizations for the same LBS log, each with a different utility to the consumer of the anonymized log. The problem of finding the optimal TP-aware anonymization is investigated. We show that trajectory-awareness renders the problem computationally harder than the trajectory-unaware variants found in the literature (NP-complete in the size of the log, versus PTIME). We describe a PTIME l-approximation algorithm for trajectories of length l and empirically show that it scales to large LBS logs (up to 2 million users).