Search for: All records

Existing studies have demonstrated that using traditional machine learning techniques, phishing detection simply based on the features of URLs can be very effective. In this paper, we explore the deep learning approach and build four RNN (Recurrent Neural Network) models that only use lexical features of URLs for detecting phishing attacks. We collect 1.5 million URLs as the dataset and show that our RNN models can achieve a higher than 99% detection accuracy without the need of any expert knowledge to manually identify the features. However, it is well known that RNNs and other deep learning techniques are still largely in black boxes. Understanding the internals of deep learning models is important and highly desirable to the improvement and proper application of the models. Therefore, in this work, we further develop several unique visualization techniques to intensively interpret how RNN models work internally in achieving the outstanding phishing detection performance. Especially, we identify and answer six important research questions, showing that our four RNN models (1) are complementary to each other and can be combined into an ensemble model with even better accuracy, (2) can well capture the relevant features that were manually extracted and used in the traditional machine learning approach for phishing detection, and (3) can help identify useful new features to enhance the accuracy of the traditional machine learning approach. Our techniques and experience in this work could be helpful for researchers to effectively apply deep learning techniques in addressing other real-world security or privacy problems. 

Context. The magnetic field is the underlying cause of solar activities. Spectropolarimetric Stokes inversions have been routinely used to extract the vector magnetic field from observations for about 40 years. In contrast, the photospheric continuum images have an observational history of more than 100 years. Aims. We suggest a new method to quickly estimate the unsigned radial component of the magnetic field, | B r |, and the transverse field, B t , just from photospheric continuum images ( I ) using deep convolutional neural networks (CNN). Methods. Two independent models, that is, I versus | B r | and I versus B t , are trained by the CNN with a residual architecture. A total of 7800 sets of data ( I , B r and B t ) covering 17 active region patches from 2011 to 2015 from the Helioseismic and Magnetic Imager are used to train and validate the models. Results. The CNN models can successfully estimate | B r | as well as B t maps in sunspot umbra, penumbra, pore, and strong network regions based on the evaluation of four active regions (test datasets). From a series of continuum images, we can also detect the emergence of a transverse magnetic field quantitatively with the trained CNN model. The three-day evolution of the averaged value of the estimated | B r | and B t from continuum images follows that from Stokes inversions well. Furthermore, our models can reproduce the nonlinear relationships between I and | B r | as well as B t , explaining why we can estimate these relationships just from continuum images. Conclusions. Our method provides an effective way to quickly estimate | B r | and B t maps from photospheric continuum images. The method can be applied to the reconstruction of the historical magnetic fields and to future observations for providing the quick look data of the magnetic fields. 

Abstract Sensing is an essential part in autonomous driving and intelligent transportation systems. It enables the vehicle to better understand itself and its surrounding environment. Vehicular networks support information sharing among different vehicles and hence enable the multi‐vehicle multi‐sensor cooperative sensing, which can greatly improve the sensing performance. However, there are a couple of issues to be addressed. First, the multi‐sensor data fusion needs to deal with heterogeneous data formats. Second, the cooperative sensing process needs to deal with low data quality and perception blind spots for some vehicles. In order to solve the above problems, in this paper the occupancy grid map is adopted to facilitate the fusion of multi‐vehicle and multi‐sensor data. The dynamic target detection frame and pixel information of the camera data are mapped to the static environment of the LiDAR point cloud, and the space‐based occupancy probability distribution kernel density estimation characterization fusion data is designed , and the occupancy grid map based on the probability level and the spatial level is generated. Real‐world experiments show that the proposed fusion framework is better compatible with the data information of different sensors and expands the sensing range by involving the collaborations among multiple vehicles in vehicular networks.