Microsoft Technology Licensing, LLC

Christian Seifert

Jack Wilson Stokes, Iii

Benjamin Zorn

Benjamin Livshits

David Earl Heckerman

Craig Henry Wittenberg

Robert Alexander Sim

Jason Joseph Weber

Jack W Stokes

Anthony Paul Penta

Robert L McCann

Elliott Jeb Haber

Sasi Parthasarathy

Tomasz Kasperkiewicz

Timothy William Burrell

Charles Strempler

Abhishek Sagar

Jonathan Bar Or

Mara Beth Fortini

Ali Alabbas

Farid Tajaddodianfar

Charles M Curtsinger

Gil Lapid Shafriri

Christina Colcernian

Long Lu

Robert G Lefferts

Christian Stockwell

Michael David Mccormack

Eric Paul Douglas

Sinclaire Renee Hamilton

Brian Keith Catlin

Richard Joseph Murillo

Himanshu Chandola

Navaneethan Santhanam

Clemens Kolbitsch

Andrew W Wicker

Peter A Cap

Kristian Holsheimer

Rupo Zhang

Paresh Maisuria

Talha Ongun

The technology described herein identifies and mitigates phishing attempts by analyzing user input received at the operating system level. Initially, a credential, such as a username or password, is registered with the threat detection system. The technology described herein intercepts user input at the operating system level, generates a hash of the input, and compares it with a hash of a credential being monitored. A credential entry is detected when a hash for the character string entered matches a hash for a credential being monitored. The technology described herein will perform a threat assessment when a secret entry is detected. The threat assessment may use the application context and the network context as inputs to the assessment. Various mitigation actions may be taken when a threat is detected.

Credential input detection and threat analysis

The techniques disclosed herein enable systems to train a machine learning model to classify malicious command line strings and select anomalous and uncertain samples for analysis. To train the machine learning model, a system receives a labeled data set containing command line inputs that are known to be malicious or benign. Utilizing a term embedding model, the system can generate aggregated numerical representations of the command line inputs for analysis by the machine learning model. The aggregated numerical representations can include various information such as term scores that represent a probability that an individual term of the command line string is malicious as well as numerical representations of the individual terms. The system can subsequently provide the aggregated numerical representations to the machine learning model for analysis. Based on the aggregated numerical representations, the machine learning model can learn to distinguish malicious command line inputs from benign inputs.

Command classification using active learning

Various embodiments discussed herein enable the detection of malicious content. Some embodiments do this by determining a similarity score between content, computer objects, or indications (e.g., vectors, file hashes, file signatures, code, etc.) known to be malicious and other content (e.g., unknown files) or indications based on feature weighting. Over various training stages, certain feature characteristics for each labeled malicious content or indication can be learned. For example, for a first malware family of computer objects, the most prominent feature may be a particular URL, whereas other features change considerably for different iterations of the first malware family of computer objects. Consequently, the particular URL can be weighted to determine a particular output classification corresponding to malicious behavior.

Detecting unknown malicious content in computer systems

Described are examples for curating threat intelligence data including receiving threat intelligence data comprising a list of entities, one or more associations between entities, a reputation score for each entity, and/or a confidence value corresponding to the one or more associations. An updated reputation score for at least one of a first type of entities can be determined based at least in part on the confidence value and/or on determining a reputation score of at least one of a second type of entities to which the at least one of the first type of entities is associated in the one or more associations. The reputation score of the at least one of the first type of entities can be updated, in the threat intelligence data, to the updated reputation score.

Techniques for curating threat intelligence data

The disclosed embodiments generate a plurality of anomaly detector configurations and compare results generated by these anomaly detectors to a reference result set. The reference result set is generated by a trained model. A correlation between each result generated by the anomaly detectors and the result set is compared to select an anomaly detector configuration that provides results most similar to those of the trained model. In some embodiments, data defining the selected configuration is then communicated to a product installation. The product installation instantiates the defined anomaly detector and analyzes local events using the instantiated detector. In some other embodiments, the defined anomaly detector is instantiated by the same system that selects the anomaly detector, and thus in these embodiments, the anomaly detector configuration is not transmitted from one system to another.

Dynamic configuration of anomaly detection

Identify a set or session of processes as having certain characteristics. A method obtains a known set or session of processes, wherein the known set or session of processes has the certain characteristics. A set or session of processes to be evaluated is obtained. A weighted similarity measure is performed between the known set or session of processes and the set or session of processes to be evaluated. The weighted similarity measure is performed element wise, where a comparison is performed for each defined element in the set or session of processes to be evaluated against elements in the known set or session of processes.

Large scale malicious process detection

One or more techniques and/or systems are provided for file acquisition for reputation evaluation. A reputation service may be configured to evaluate files and provide reputations of such files to clients (e.g., an indication as to whether a file is safe or malicious). If the reputation service receives a reputation request for a file that is unknown to the reputation service (e.g., a file not yet fully acquired by the reputation service), then the reputation service may identify a set of chunks into which the file can be partitioned. The reputation service may obtain chunks from various clients, such as a first chunk from a first client and a second chunk from a second client. Such chunks may be evaluated to assign a reputation to the file. In this way, the reputation service may retrieve portions of a file in a distributed manner for reputation evaluation.

Chunk-based file acquisition and file reputation evaluation

A machine-implemented method for detecting scareware includes the steps of accessing one or more landing pages to be evaluated, extracting one or more features from the landing pages, and providing a classifier to compare the features extracted from the landing pages with features of known scareware and non-scareware pages. The classifier determines a likelihood that the landing page is scareware. If determined to be scareware, the landing page is removed from search results generated by a search engine. The features can be URLs, text, image interest points, image descriptors, a number of pop-ups generated, IP addresses, hostnames, domain names, text derived from images, images, metadata, identifiers of executables, and combinations thereof.

Scareware detection

Techniques for execution of multiple execution paths are described. In one or more embodiments, an execution of a portion of executable code is conditioned upon a particular environment-specific value. For example, the execution of the executable code can cause one type of output if the value of the variable equals a particular value, and can cause a different type of output if the value of the variable equals a different value. Techniques discussed herein can enable the executable code to be executed such that multiple outputs are produced, e.g., by executing the code according to the different values for the variable. In implementations, the multiple outputs can be analyzed for various attributes, such as presence of malware, implementation and coding errors, and so on.

Execution of multiple execution paths

This document describes techniques for detection of code-based malware. According to some embodiments, the techniques utilize a collection of known malicious code and know benign code and determine which features of each type of code can be used to determine whether unclassified code is malicious or benign. The features can then be used to train a classifier (e.g., a Bayesian classifier) to characterize unclassified code as malicious or benign. In at least some embodiments, the techniques can be used as part of and/or in cooperation with a web browser to inspect web content (e.g., a web page) to determine if the content includes code-based malware.

Detection of code-based malware

Growing community of inventors

Seattle, WA, United States of America

Christian Seifert