arXiv subjects
Publications and source records attributed to Denis Turdakov.
This paper aims to motivate stochastic optimization problems from a statistical perspective and a statistical learning perspective, where the goal is to maximize the log-likelihood or minimize the population risk. We briefly describe the two main approaches: offline (Monte Carlo / Sample Average Approximation) and online (Stochastic Approximation) approaches -- to solve the expectation minimization problems.
The vulnerability of artificial neural networks to adversarial perturbations in the black-box setting is widely studied in the literature. The majority of attack methods to construct these perturbations suffer from an impractically large number of queries required to find an adversarial example. In this work, we focus on knowledge distillation as an approach to conduct transfer-based black-box adversarial attacks and propose an iterative training of the surrogate model on an expanding dataset. This work is the first, to our knowledge, to provide provable guarantees on the success of knowledge distillation-based attack on classification neural networks: we prove that if the student model has enough learning capabilities, the attack on the teacher model is guaranteed to be found within the finite number of distillation iterations.
Online network crawling tasks require a lot of efforts for the researchers to collect the data. One of them is identification of important nodes, which has many applications starting from viral marketing to the prevention of disease spread. Various crawling algorithms has been suggested but their efficiency is not studied well. In this paper we compared six known crawlers on the task of collecting the fraction of the most influential nodes of graph. We analyzed crawlers behavior for four measures of node influence: node degree, k-coreness, betweenness centrality, and eccentricity. The experiments confirmed that greedy methods perform the best in many settings, but the cases exist when they are very inefficient.
Random graph (RG) models play a central role in the complex networks analysis. They help to understand, control, and predict phenomena occurring, for instance, in social networks, biological networks, the Internet, etc. Despite a large number of RG models presented in the literature, there are few concepts underlying them. Instead of trying to classify a wide variety of very dispersed models, we capture and describe concepts they exploit considering preferential attachment, copying principle, hyperbolic geometry, recursively defined structure, edge switching, Monte Carlo sampling, etc. We analyze RG models, extract their basic principles, and build a taxonomy of concepts they are based on. We also discuss how these concepts are combined in RG models and how they work in typical applications like benchmarks, null models, and data anonymization.
Social networks crawling is in the focus of active research the last years. One of the challenging task is to collect target nodes in an initially unknown graph given a budget of crawling steps. Predicting a node property based on its partially known neighbourhood is at the heart of a successful crawler. In this paper we adopt graph neural networks for this purpose and show they are competitive to traditional classifiers and are better for individual cases. Additionally we suggest a training sample boosting technique, which helps to diversify the training set at early stages of crawling and thus improves the predictor quality. The experimental study on three types of target set topology indicates GNN based approach has a potential in crawling task, especially in the case of distributed target nodes.