with WHOIS XML API
and Networking (ICAR-CNR); DIMES, University of Calabria
Malicious URL Detection via Machine Learning
Protection against malicious websites is an important task in cybersecurity. A common way of identifying such sites is the use of blacklists which contain a large set of URLs considered dangerous. There are various techniques for compiling such lists, and there is obviously a need for methods to verify if a suspicious site is really dangerous. Also, as the number of malicious sites is extremely high and changes frequently over time, it is virtually impossible to create perfect blacklists. Hence, methods for identifying malicious sites are of significant practical importance and are a subject of relevant scientific interest, too.
A machine learning approach
When a site is suspected to be malicious, WHOIS data can serve as a good basis for an investigation. In Ref.  a machine-learning approach was introduced. The idea was to use WHOIS information available on a given set of URLs to design a classifier to decide if a site is malicious.
The idea is to consider features of URLs. These include lexicographical information such as, e.g. the length of an URL, or whether it contains the word “Login“ which frequently appears in malicious ones. Another set of features comes from WHOIS data, as malicious URLs frequently come from certain geolocations, they exist for a very limited time only, etc.
Mathematically these features can be modeled as directions in a multi-dimensional space, i.e. a dimension for each feature set. In this space each URL is represented by a point. The space is then split into two parts, one contains most of the malicious URLs, while the points of benigns are in the other part. Supervised learning consists in finding the geometrical object separating the two parts as accurately as possible, using a large “training set“ of URLs, already labeled as malicious/non-malicious. If this is done, one has a classifier at hand: an arbitrary URL can be identified as malicious or benign with some accuracy, depending on the part of the hyperspace in which its point resides.
In Ref.  the subsets are assumed to be separated by a hyperplane, the multi-dimensional version of a plane. The finding of this plane is a large-scale optimization problem.
Recently Astorino et al. have published  a method in the same spirit. Their methodology is aimed to be suitable for very large datasets, too. In addition, spherical separation, in which a multidimensional spherical surface is considered, is utilized as it appears to have benefits over hyperplanes. They consider a limited number of features in order to prevent the explosion of the size of the sample space and have adopted a low complexity algorithm which possibly yields a somewhat less accurate characterization, but is computationally efficient and thus suitable for large datasets.
In Ref.  the details of the methods are described and their successful operation is demonstrated on various real-life datasets. The biggest considered dataset consists of 11,975 URLs, 5,090 malicious and 6,885 benign.
The research was carried out by a research group of University of Calabria, Italy. The implementation of the method needed a large amount of accurate WHOIS data. Queries had to be implemented in the programming languages preferred by the researchers and had to run quickly to obtain the relevant information. This was realized using the services of WhoisXML API.
 Ma J, Saul LK, Savage S, and Voelker GM. Beyond
blacklists: learning to detect malicious web sites from suspicious URLs. KDD'09, pages 1245-1253, Paris,
France, June 28 July 1 2009.
 A. Astorino, A. Chiarello, M. Gaudioso, and A. Piccolo.
Malicious URL detection via spherical classification. Neural Computing and Applications, Jun 2016.
We work hard to improve our services for you. As part of that, we welcome your feedback, questions and suggestions. Please let us know your thoughts and feelings, and any way in which you think we can improve our product.
For a quick response, please select the request type that best suits your needs. For more info regarding the request types, see the Contact us page.