Microsoft Technology Licensing, LLC

Andrey Karpovsky

Roy Levin

Tomer Rotstein

Tamer Salman

Yotam Livny

Ram Haim Pliskin

Naama Kraus

Ron Matchoro

Michael Makhlevich

Fady Nasereldeen

Shay Chriba Sakazi

Moshe Israel

Idan Yehoshua Hen

Evgeny Bogokovsky

Ariel Brukman

Yoav M Frandzel

Fady Naser El Deen

Haim Saadia Bendanan

Coral Cohen

Josef Weizman

Hani Hana Neuvirth

Tomer Weisberg

Oren Yossef

Dawn Antonette Burns

Shalom Shay Shavit

Ram Bracha

Tomer Levav

Amit Magen Medina

Arieh Bibliowicz

Michael Kuritzky Bakman

Yoav Y Rubin

Shimon Ezra

Omer Savion

Inbal Argov

Asaf Nakash

Moran Polack

Anastasia Kovalkov

Haim Saadia Ben Danan

A computing system is configured to train a machine-learning model for detecting suspicious network activities based on a training dataset. The training of the machine-learning model may be supervised or unsupervised training. The training dataset includes multiple strings. For each of the multiple strings, the computing system extracts one or more N-grams substrings, where N is a natural number that is equal to or greater than 2. The computing system then determines a probability of each N-grams substring that may occur in a string. When the machine-learning model is executed, it is configured to classify whether a given string contained in network communication is a random string. In response to classifying that the given string is a random string, an alert is generated at a particular computing system to which the network communication is directed.

Random strings classification for detecting suspicious network activities

Techniques are described herein that are capable of performing a security action regarding an access request using a cluster-based determination of malicious intent. Subsets of access requests are clustered into respective clusters, which correspond to respective requestor types, based at least on the access requests in the subsets having respective attributes that indicate the respective requestor types. A determination is made that a received access request, which is associated with a designated requestor type, is malicious based at least on a relationship between an access action, which is requested by the received access request using an access token, and an access behavior associated with a group of access requests in a designated cluster that use the access token. As a result of determining that the received access request is malicious, a security action is performed with regard to the received access request.

Performing a security action regarding an access request using a cluster-based determination of malicious intent

Techniques are described herein that are capable of performing a security action with regard to an access token based on clustering of access requests. Subsets of access requests are clustered into respective clusters, which correspond to respective requestor types, based at least on the access requests in the subsets having respective attributes that indicate the respective requestor types. The access requests request access to cloud resources. Access behavior(s) associated with the access requests that are included in respective cluster(s) are identified. A security action is performed with regard to an access token based at least on at least one of the access behavior(s).

Performing a security action with regard to an access token based on clustering of access requests

The techniques disclosed herein prevent a rogue resource from being created within a cloud computing environment. For example, a rogue serverless function may be prevented from integrating with a cloud-based database, thereby preventing the serverless function from performing malicious operations such as low-rate data exfiltration. The rogue serverless function is detected before it is installed, heading off the attack completely. In some configurations, a key retrieval request is received. Parameters of the key retrieval request are analyzed for anomalies, and anomalous key retrieval requests are stored in a pool. Then, when a request to create a resource is received, the pool of anomalous key retrieval requests is searched for a match. When a match is found, the resource creation request may be suspended pending a further security review.

Detecting suspicious data access by a rogue cloud resource

A computer implemented method includes obtaining a list of information storage resources that contain electronic files and forming multiple clusters of the information storage resources from the list of information storage resources. A subset of the information storage resources is selected from the clusters. Each subset the information storage resources is scanned for a selected type of content. The selected information storage resources are flagged as containing the selected type of content as a function of the scanning finding the selected type of content.

Smart sampling for sensitive data detection

Methods, systems, apparatuses, and computer-readable storage mediums described herein are configured to detect mass control plane operations, which may be indicative of anomalous (or malicious) behavior. For example, one or more logs that specify a plurality of access enablement operations performed with respect to an entity is received. The log(s) are analyzed to identify a number of access enablement operations that occurred in a particular time period. A determination is made as to whether the identified number of access enablement operations meets a threshold condition (e.g., to determine whether an unusually high number of such operations occurred in a given time period). Based on the threshold condition being met, a determination is made that anomalous behavior has occurred with respect to the entity. Responsive to determining that the potentially behavior has occurred, a mitigation action may be performed that mitigates the behavior.

Detecting mass control plane operations

A system for dynamically determining the legitimacy of a source internet protocol (IP) address requesting access to a target resource includes an address classifier, a resource similarity identifier, and a connection legitimacy prediction engine. The IP address classifier classifies the source IP address into a relevant address group selected from among a plurality of address groups. Each of the address groups consist of IP addresses that satisfy at least one address similarity criterion. The resource similarity identifier identifies a group of similar resources for the target resource based commonalities in a first subset of the address groups that have previously accessed the target resource and subsets of the address groups that have accessed each of the similar resources. The connection legitimacy prediction engine compute a probability of receiving a legitimate new request to access the target resource from the relevant address group based on connection history data associated with the relevant address group and each of the similar resources. The system denies the request to access the target resource in response to determining that the probability falls below a defined threshold.

Intelligent auto-detection of anomalous web-based access requests

According to examples, an apparatus may include a processor that may calculate a normalized threat intelligence score (TIS) for an autonomous system (AS) based on a sum of threat intelligence (TI) signals associated with Internet protocol (IP) addresses controlled by the AS and a count of the IP addresses controlled by the AS. The processor may also determine, based on the normalized TIS for the AS, a probability that activities associated with the IP addresses controlled by the AS are likely to be malicious. The processor may further output the determined probability that the activities associated with the IP addresses controlled by the AS are likely to be malicious.

Malicious activity probability determinations for autonomous systems

A computing system generates from received user input an initial profile. The initial profile specifies expected behavioral patterns of datasets that are to be received by the computing system. The computing system extracts from received datasets features that are indicative of behavioral patterns of the received datasets. The computing system provides the initial profile to first machine-learning models. The first machine-learning models have been trained using a subset of the received datasets. The first machine-learning models use the initial profile to determine if the behavioral patterns of the received datasets are anomalous. The computing system includes second machine-learning models that have been trained using a subset of the received datasets. The second machine-learning models train a second profile based on the extracted features to specify behavioral patterns of the received datasets that are learned by the second machine-learning model.

Machine learning approach for solving the cold start problem in stateful models

Techniques are described herein that are capable of detecting malicious obfuscation in a SQL statement based at least in part on an effect and/or processed version of the SQL statement. In a first example, a raw version of a SQL statement is compared to a processed version of the SQL statement. A determination is made that a command in the processed version is not included in the raw version. The raw version is detected to be malicious based at least in part on the determination. In a second example, a SQL statement is bound to an event that results from execution of the SQL statement. Textual content of the SQL statement and an effect of the event are compared. The SQL statement is detected to be malicious based at least in part on the effect of the event not being indicated by the textual content.

Detecting malicious obfuscation in a SQL statement based on an effect and/or processed version thereof

Methods, systems, apparatuses, and computer-readable storage mediums are described for adapting a spike detection algorithm. A first detection algorithm that monitors a first set of events in a computing environment is executed. A set of constraint metrics in the computing environment are monitored. Based on the monitored set of constraint metrics, a second detection algorithm is generated. The second detection algorithm is an adapted version of the first detection algorithm and is configured to monitor a second set of events in the computing environment. The second detection algorithm is executed, and a remediation action is performed in response to an abnormal event detected in the computing environment by the second detection algorithm.

Adaptable framework for spike detection under dynamic constraints

Techniques are described herein that are capable of using graph enrichment to detect a potentially malicious access attempt. A graph that includes nodes and configuration-based links is generated. The nodes represent respective resources. Behavior-based links are added to the graph based at least in part on traffic logs associated with at least a subset of the resources. An attempt to create a new behavior-based link is identified. A probability of the new behavior-based link being created in the graph is determined. The probability is based at least in part on the configuration-based links and the behavior-based links. The new behavior-based link is identified as a potentially malicious link based at least in part on the probability being less than or equal to a threshold probability. A security action is performed based at least in part on the new behavior-based link being identified as a potentially malicious link.

Using graph enrichment to detect a potentially malicious access attempt

The processing of an incoming query targeted to a data store in a way that early detection of code injections can be achieved. Initial code injections, even if unsuccessful, can be used to adjust the code injections to more successfully perform harmful actions against the data store. Accordingly, early detection can be used to block attackers from experimenting against the data store. The early detection is accomplished by detecting when all or a portion of the query is structured in accordance with a query language, but does not follow the syntax recognized by the underlying data store. This is a good indication that the issuer of the query is performing a blind code injection, not knowing the type of the underlying data store.

Code injection detection using syntactic deviation

The detection and alerting on malicious queries that are directed towards a data store. The detection is done by using syntax metrics of the query. This can be done without evaluating (or at least without retaining) the unmasked query. In order to detect a potentially malicious query, syntax metric(s) of that query are accessed. The syntax metric(s) are then fed into a model that is configured to predict maliciousness of the query based on the one or more syntax metrics. The output of the model then represents a prediction of maliciousness of the query. Based on the output of the model representing the predicted maliciousness, a computing entity associated with the data store is then alerted.

Detecting malicious queries using syntax metrics

Multiple variate anomaly detection against multiple scopes of the requested resource. Even if one of the variates patterns is sensitive to physical location of the requestor and/or the resource, not all of the variates of the access pattern will be. Furthermore, even if one of the scopes of the resource is sensitive to physical location of the resource, not all scopes will be. Thus, the use of multiple variates of the access pattern and multiple scopes of the anomaly detection allows for better estimates of anomaly detection to be made, even when the source of the access request is virtualized and/or the location of the resource is virtualized.

Growing community of inventors

Kiryat Motzkin, Israel

Andrey Karpovsky