International Business Machines Corporation

Andrey Finkelshtein

Eitan Menahem

Noga Agmon

Bar Haim

Yehonatan Bitton

Oded Margalit

Itay Hazan

Anton Puzanov

Shimon Benusovich

Sarit Hollander

Yuval Margalit

Mechanisms are provided for detecting fraudulent user flows associated with a website. User flow data, representing an interaction by a user with content of a website, is received and converted to a vector representation that represents a time series transition from one portion of website content to another of the website. The vector representation is input to a trained sequential machine learning computer model which generates a classification of the vector representation. A determination as to whether or not the user flow data represents a fraudulent user flow is made based on the classification. An output is generated that indicates whether or not the user flow is a fraudulent user flow based on the detection.

Detecting fraudulent user flows

A system and a method for training and classification using an optimized classification schema using an ensemble of cascaded classifiers is disclosed. Each of the cascaded classifiers is characterized by a set of classifier parameters and the classifiers which are not the first in a cascade are associated with one or more thresholds used to determine when to execute them according to a confidence measure computed by a preceding cascaded classifier. The optimization comprises a genetic algorithm applied to a set of ensembles of classification and parameters and the set of scores, into a pool of ensembles and associated scores. The scores may be based on associated classification quality and cost.

Optimizing cascade of classifiers schema using genetic search

An example system includes a processor to receive mouse event data of a session. The processor is to split the mouse event data of the session into mouse event n-grams. The processor is to extract features from the mouse event n-grams. The processor is to send the extracted features to a trained machine learning model. The processor is to receive an output decision from the trained machine learning model.

Machine learning model trained using features extracted from n-grams of mouse event data

A cyber security method including: obtaining user flow data associated with a browsing session at a website; constructing a directed graph representative of the browsing session; computing a set of features for the directed graph; and applying a machine learning classifier to the set of features, to classify the browsing session as legitimate or fraudulent.

User flow graph analytics for cyber security

A machine learning clustering process is trained. Web pages of a website are clustered. User flow data associated with a first browsing session at the website is obtained. The user flow data includes a plurality of web page identifiers (e.g., URLs). A web page record for each of the web page identifiers is generated. Each web page record includes words of the corresponding web page identifier. Clusters of web page identifiers previously output from the trained machine learning clustering process are received. For each of the web page records, a cluster of web page identifiers is identified by mapping the web page record to one of the clusters of web page identifiers using the machine learning clustering process. A directed graph representative of the first browsing session is constructed. One or more nodes of the directed graph are the identified clusters of web page identifiers.

Clustering web page addresses for website analysis

Automated detection of personal information in free text, which includes: automatically applying a named-entity recognition (NER) algorithm to a digital text document, to detect named entities appearing in the digital text document, wherein the named entities are selected from the group consisting of: at least one person-type entity, and at least one non-person-type entity; automatically detecting at least one relation between the named entities, by applying a parts-of-speech (POS) tagging algorithm and a dependency parsing algorithm to sentences of the digital text document which contain the detected named entities; automatically estimating whether the at least one relation between the named entities is indicative of personal information; and automatically issuing a notification of a result of the estimation.

Automated detection of personal information in free text

A method for training thresholds controlling data flow in a plurality of cascaded classifiers for classifying malicious software, comprising: in each of a plurality of iterations: computing a set of scores, each for one of a set of threshold sequences, each threshold sequence is a sequence of sets of classifier output thresholds, each set of classifier output thresholds used to control a flow of data from a first cascaded classifier of the plurality of cascaded classifiers to a second cascaded classifier of the plurality of cascaded classifiers, each score computed when classifying, using the respective threshold sequence, each of a plurality of software objects as one of a set of maliciousness classes; computing a set of new threshold sequences by applying a genetic algorithm to the set of threshold sequences and the set of scores; and using the set of new threshold sequences in a consecutive iteration.

System and method for staged ensemble classification

A method for quantifying a similarity between a target dataset and multiple source datasets and identifying one or more source datasets that are most similar to the target dataset is provided. The method includes receiving, at a computing system, source datasets relating to a source domain and a target dataset relating to a target domain of interest. Each dataset is arranged in a tabular format including columns and rows, and the source datasets and the target dataset include a same feature space. The method also includes pre-processing, via a processor of the computing system, each source-target dataset pair to remove non-intersecting columns. The method further includes calculating at least two of a dataset similarity score, a row similarity score, and a column similarity score for each source-target dataset pair, and summarizing the calculated similarity scores to identify one or more source datasets that are most similar to the target dataset.

Identifying source datasets that fit a transfer learning process for a target domain

An example system includes a processor to receive a current session and previous sessions associated with an account. The processor can split the current session and the previous sessions into action windows. The processor can calculate a window similarity score for each action window of the current session using a pair-wise comparison with action windows of each of the previous sessions. The processor can aggregate the window similarity scores to generate a replay likelihood score for the current session with respect to each of the previous sessions. The processor can classify the current session as a replay attack in response to detecting that a replay likelihood score of the current session exceeds a threshold.

Detecting replay attacks using action windows

Embodiments of the present systems and methods may decide if a software file is malicious or benign, using properties of the file's overlay, if existing. For example, in an embodiment, a computer-implemented method for identifying malware in computer systems may comprise receiving a plurality of executable files labeled as being malicious or benign, training a machine learning model using properties extracted from overlays associated with each of the plurality of received labeled executable files, receiving an executable file that is not labeled, determining whether the received unlabeled executable file is malicious or benign using the trained machine learning model based on properties extracted from an overlay associated with the received unlabeled executable file, and transmitting information identifying the received unlabeled executable file as malicious when the received unlabeled executable file is determined to be malicious.

Growing community of inventors

Beer Sheva, Israel

Andrey Finkelshtein