Abstract Tor is an open source software that allows accessing various kinds of resources, known as hidden services, while guaranteeing sender and receiver anonymity. Tor relies on a free, worldwide, overlay network, managed by volunteers, that works according to the principles of onion routing in which messages are encapsulated in layers of encryption, analogous to layers of an onion. The Tor Web is the set of web resources that exist on the Tor network, and Tor websites are part of the so-called dark web. Recent research works have evaluated Tor security, its evolution over time, and its thematic organization. Nevertheless, limited information is available about the structure of the graph defined by the network of Tor websites, not to be mistaken with the network of nodes that supports the onion routing. The limited number of entry points that can be used to crawl the network, makes the study of this graph far from being simple. In the present paper we analyze two graph representations of the Tor Web and the relationship between contents and structural features, considering three crawling datasets collected over a five-month time frame. Among other findings, we show that Tor consists of a tiny strongly connected component, in which link directories play a central role, and of a multitude of services that can (only) be reached from there. From this viewpoint, the graph appears inefficient. Nevertheless, if we only consider mutual connections, a more efficient subgraph emerges, that is, probably, the backbone of social interactions in Tor.

Efficient HPC libraries often expose multiple tunable parameters, algorithmic implementations, or a combination of them, to provide optimized routines. The optimal parameters and algorithmic choices may depend on input properties such as the shapes of the matrices involved in the operation. Traditionally, these parameters are manually tuned or set by auto-tuners. In emerging applications such as deep learning, this approach is not effective across the wide range of inputs and architectures used in practice. In this work, we analyze different machine learning techniques and predictive models to accelerate the convolution operator and GEMM. Moreover, we address the problem of dataset generation, and we study the performance, accuracy, and generalization ability of the models. Our insights allow us to improve the performance of computationally expensive deep learning primitives on high-end GPUs as well as low-power embedded GPU architectures on three different libraries. Experimental results show significant improvement in the target applications from 50% up to 300% compared to auto-tuned and high-optimized vendor-based heuristics by using simple decision tree- and MLP-based models.

Network-based ranking methods (e.g., centrality analysis) have found extensive use in systems biology and network medicine for the prediction of essential proteins, for the prioritization of drug targets candidates in the treatment of several pathologies and in biomarker discovery, and for human disease genes identification. We here studied the connectivity of the human protein-protein interaction network (i.e., the interactome) to find the nodes whose removal has the heaviest impact on the network, i.e., maximizes its fragmentation. Such nodes are known as Critical Nodes (CNs). Specifically, we implemented a Critical Node Heuristic (CNH) and compared its performance against other four heuristics based on well known centrality measures. To better understand the structure of the interactome, the CNs' role played in the network, and the different heuristics' capabilities to grasp biologically relevant nodes, we compared the sets of nodes identified as CNs by each heuristic with two experimentally validated sets of essential genes, i.e., the genes whose removal impact on a given organism's ability to survive. Our results show that classical centrality measures (i.e., closeness centrality, degree) found more essential genes with respect to CNH on the current version of the human interactome, however the removal of such nodes does not have the greatest impact on interactome connectivity, while, interestingly, the genes identified by CNH show peculiar characteristics both from the topological and the biological point of view. Finally, even if a relevant fraction of essential genes is found via the classical centrality measures, the same measures seem to fail in identifying the whole set of essential genes, suggesting once again that some of them are not central in the network, that there may be biases in the current interaction data, and that different, combined graph theoretical and other techniques should be applied for their discovery.