Abstract: A machine learning model is trained using tuples that identify an actor, a resource, and a rating based on a normalized count of the actor's attempts to access the resource. Actors may be users, groups, IP addresses, or otherwise defined. Resources may be storage, virtual machines, APIs, or otherwise defined. A risk assessor code feeds an actor-resource pair to the trained model, which computes a recommendation score using collaborative filtering. The risk assessor inverts the recommendation score to obtain a risk measurement; a low recommendation score corresponds to a high risk, and vice versa. The risk assessor code or other code takes cybersecurity action based on the recommendation score. Code may accept a risk R, or aid mitigation of the risk R, where R denotes a risk that the scored pair represents an unauthorized attempt by the pair actor to access the pair resource.

Abstract: An indication is received of a security alert. The indication is generated based on a detected anomaly in one of a data plane or a control plane of a computing environment. When the detected anomaly is in the data plane, the control plane is monitored for a subsequent anomaly in the control plane, and otherwise the data plane is monitored for a subsequent anomaly in the data plane. A correlation between the detected anomalies is determined. A notification of the security alert is sent when the correlation exceeds a predetermined threshold.

Abstract: Enriched access data supports anomaly detection to enhance network cybersecurity. Network access data is enriched using service nodes representing resource provision and other services, with geolocation nodes representing grouped access origins, and access values representing access legitimacy confidence. Data enrichment provides a trained model by mapping IP addresses to geolocations, building a bipartite access graph whose inter-node links indicate aspects of accesses from geolocations to services, and generating semantic vectors from the graph. Vector generation may include collaborative filtering, autoencoding, neural net embedding, and other machine learning tools and techniques. Anomaly detection systems then calculate service-geolocation or geolocation-geolocation vector distances with anomaly candidate vectors and the model's graph-based vectors, and treat distances past a threshold as anomaly indicators.

Abstract: Methods, systems and computer program products are provided for detection of hacker tools based on their network signatures. A suspicious process detector (SPD) may be implemented on local computing devices or on servers to identify suspicious (e.g., potentially malicious) or malicious executables. An SPD may detect suspicious and/or malicious executables based on the network signatures they generate when executed as processes. An SPD may include a model, which may be trained based on network signatures generated by multiple processes on multiple computing devices. Computing devices may log information about network events, including the process that generated each network event. Network activity logs may record the network signatures of one or more processes. Network signatures may be used to train a model for a local and/or server-based SPD. Network signatures may be provided to an SPD to detect suspicious or malicious executables using a trained model.

Abstract: Generally discussed herein are devices, systems, and methods for cloud traffic monitoring. A method can include receiving sampled network metadata of a packet transmitted via a computer network, providing the sampled network metadata to a neural network (NN) trained on labeled sampled network metadata, and providing, based on only the sampled network metadata, a classification for the sampled network metadata via the trained neural network.

Abstract: Systems, methods, and apparatuses are provided for clustering incidents in a computing environment. An incident notification relating to an event (e.g., a potential cyberthreat or any other alert) in the computing environment is received and a set of features may be generated based on the incident notification. The set of features may be provided as an input to a machine-learning engine to identify a similar incident notification in the computing environment. The similar incident notification may include a resolved incident notification or an unresolved incident notification. An action to resolve the incident notification may be received, and the received action may thereby be executed. In some implementations, in addition to resolving the received incident notification, the action may be executed to resolve a similar unresolved incident notification identified by the machine-learning engine.

Abstract: A pseudo-relevance feedback (PRF) system is disclosed that determines an optimized relevance model for a search query by utilizing a posterior relevance model to estimate the likelihood that an initial set of top-K retrieved documents would be retrieved given the posterior relevance model, re-ranking the top-K documents based on their respective estimates of likelihood of retrieval, determining a rank similarity between the initial ranking of the top-K documents and the re-ranking of the top-K documents, updating one or more model parameters of the posterior relevance model based on the rank similarity, and iteratively performing the above process until the rank similarity is maximized, at which point, the optimized relevance model is obtained.

Abstract: Methods, systems, and media are shown for generating access control rules for computer resources involving collecting historical access data for user accesses to a computer resource and separating the historical access data into a training data set and a validation data set. An access control rule is generated for the computer resource based on the properties of the user accesses to the computer resource in the training data set. The rule is validated against the validation data set to determine whether the rule produces a denial rate level is below a threshold when the rule is applied to the validation data set. If the rule is valid, then it is provided to an administrative interface so that an administrator can select the rule for application to incoming user requests.

Abstract: Unauthorized use of user credentials in a network is detected. Data indicative of text strings being used to access resources in the network is accessed. Regex models are determined for the text strings. Groupings of the regex models are determined based on an optimization of a cumulative weighted function. A regex model having a cumulative weighted function that exceeds a predetermined threshold is identified. An alert is generated when the cumulative weighted function for the identified regex model exceeds the predetermined threshold.

Abstract: Generally discussed herein are devices, systems, and methods for computer or other network device security. A method can include identifying a profile associated with event data regarding an operation performed on a cloud resource, determining whether the event data is associated with anomalous customer interaction with the cloud resource, in response to determining the event data is associated with anomalous customer interaction, identifying whether another cloud resource of the cloud resources with a lower granularity profile that is associated with the profile of the cloud resource has previously been determined to be a target of an anomalous operation, and providing a single alert to a client device indicating the anomalous behavior on the cloud resource in response to determining both the event data is associated with anomalous customer interaction and the another cloud resource is determined to be the target of the anomalous operation.

Abstract: Methods, systems, and computer program products are described herein for detecting malicious cloud-based resource allocations. Such detection may be achieved using machine learning-based techniques that analyze sequences of cloud-based resource allocations to determine whether such sequences are performed with a malicious intent. For instance, a sequence classification model may be generated by training a machine learning-based algorithm on both resource allocation sequences that are known to be used for malicious purposes and resource allocation sequences that are known to be used for non-malicious or benign purposes. Using these sequences, the machine learning-based algorithm learns what constitutes a malicious resource allocation sequence and generates the sequence classification model.

Abstract: Technology automatically groups security alerts into incidents using data about earlier groupings. A machine learning model is trained with select data about past alert-incident grouping actions. The trained model prioritizes new alerts and aids alert investigation by rapidly and accurately grouping alerts with incidents. The groupings are provided directly to an analyst or fed into a security information and event management tool. Training data may include entity identifiers, alert identifiers, incident identifiers, action indicators, action times, and optionally incident classifications. Investigative options presented to an analyst but not exercised may serve as training data. Incident updates produced by the trained model may add an alert to an incident, remove an alert, merge two incidents, divide an incident, or create an incident. Personalized incident updates may be based on a particular analyst's historic manual investigation actions.

Abstract: Described technologies enhance cybersecurity by leveraging collaborative filtering tools and techniques for security use by scoring attempts to access items in digital storage. Examples provided illustrate usage of accessor IDs and storage item IDs to compute recommendation scores which then operate as inverse measures of intrusion risk. Actions taken in response to recommendation scores that fall below a specified threshold may include preventing or terminating access, or alerting an administrator, for instance. A requested access may be allowed when the computed recommendation score is above a specified threshold, which indicates an acceptably low risk that the access is an unauthorized intrusion. Described cybersecurity technologies may be used by, or incorporated within, cloud services, cloud infrastructure, or virtual machines. Described cybersecurity technologies may also be used outside a cloud, e.g.

Abstract: Anomalous sequences are detected by approximating user sessions with heuristically extracted event sequences, allowing behavior analysis even without user identification or session identifiers. Extraction delimiters may include event count or event timing constraints. Event sequences extracted from logs or other event lists are vectorized and embedded in a vector space. A machine learning model similarity function measures anomalousness of a candidate sequence relative to a specified history, thus computing an anomaly score. Restrictions may be placed on the history to focus on a particular IP address or time frame, without retraining the model. Anomalous sequences may generate alerts, prompt investigations by security personnel, trigger automatic mitigation, trigger automatic acceptance, trigger tool configuration actions, or result in other cybersecurity actions.

Abstract: Cybersecurity anomaly explainability is enhanced, with particular attention to collaborative filter-based anomaly detection. An enhanced system obtains user behavior vectors derived from a trained collaborative filter, computes a similarity measure of user behavior based on a distance between user behavior vectors and a similarity threshold, and automatically produces an explanation of a detected cybersecurity anomaly. The explanation describes a change in user behavior similarity, in human-friendly terms, such as “User X from Sales is now behaving like a network administrator.” Each user behavior vector includes latent features, and corresponds to access attempts or other behavior of a user with respect to a monitored computing system. Users may be sorted according to behavioral similarity. Explanations may associate a collaborative filter anomaly detection result with a change in behavior of an identified user or cluster of users, per specified explanation structures.

Abstract: Tools and techniques are described to automate triage of security and operational alerts. Insight instances extracted from raw event data associated with an alert are aggregated, vectorized, and assigned confidence scores through classification based on machine learning. Confidence scoring enables heavily loaded administrators and controls to focus attention and resources where they are most likely to protect or improve the functionality of a monitored system. Feature vectors receive a broad base in the underlying instance values through aggregation, even when the number of instance values is unknown prior to receipt of the event data. Visibility into the confidence scoring process may be provided, to allow tuning or inform further training of a classifier model. Performance metrics are defined, and production level performance may be achieved.

Abstract: A system for predicting vulnerability of network resources is provided. The system can calculate an initial vulnerability score for each of the network resources and use the initial vulnerability scores along with activity data of the network resources to train a vulnerability model. After training, the vulnerability model can predict the vulnerability of the network resources based on new activity data collected from the network resources. Based on the predicted vulnerability, vulnerable network resources can be identified. Further analysis can be performed by comparing the activities of the vulnerable network resources and other network resources to identify activity patterns unique to the vulnerable network resources as attack patterns. Based on the attack patterns, one or more actions can be taken to increase the security of the vulnerable network resources to avoid further vulnerability.

Abstract: Disclosed herein is a system for predicting, given a pattern of triggered alerts, a next alert in order to identify malicious activity that is about to occur on resource(s) being monitored by a security operations center. A resource can include a server, a storage device, a user device (e.g., a personal computer, a tablet computer, a smartphone, etc.), a virtual machine, networking equipment, etc. Accordingly, the next alert is speculatively triggered in advance and a security analyst can be notified of a pattern of activity that is likely to be malicious. The security analyst can then investigate the pattern of triggered alerts and the speculatively triggered alert to determine whether steps to mitigate the malicious activity before it occurs should be taken.

Abstract: Generally discussed herein are devices, systems, and methods for computer or other network device security. A method can include providing an alert to a device of a first cloud user in response to determining an operation performed on a cloud resource is inconsistent with a behavior profile that defines normal operation of the cloud resource, receiving feedback from the first cloud user regarding the alert, and generating, for a second, different cloud user and by prioritizing a second alert based on the feedback from the first cloud user, a second alert.

Abstract: An anomalous user session detector is disclosed. A sequence of operations in a logon session for an authorized user is gathered. A supervised learning model is trained to identify the authorized user from the sequence of operations. An anomalous session is detected by querying the supervised learning model.