Cisco Technology, Inc.

Jan Brabec

Lukas Machlica

Martin Kopp

Jan Kohout

Cenek Skarda

Petr Somol

Ivan Nikolaev

Lukas Bajer

Danila Khikhlukha

Radek Starosta

Marc René Jacques Marie Dupont

Tomas Komarek

Marc Dupont

In some aspects, the techniques described herein relate to a method for detecting malicious emails, the method including: receiving an email, wherein the email is associated with a markup payload; determining, based on the markup payload, text data associated with the email; determining, using the text data and a first machine learning model, a first representation of the email representing text associated with the email; rendering the email to generate image data that represents a rendering of the email; determining, using the image data and a second machine learning model, a second representation of the email that represents at least the rendering of the email; and determining a prediction for the email based on the first representation and the second representation, wherein the prediction represents whether the email is predicted to be malicious based on the first representation and the second representation.

Multi-modal models for detecting malicious emails

In one embodiment, a device obtains input features for a neural network-based model. The device pre-defines a set of neurons of the model to represent known behaviors associated with the input features. The device constrains weights for a plurality of outputs of the model. The device trains the neural network-based model using the constrained weights for the plurality of outputs of the model and by excluding the pre-defined set of neurons from updates during the training.

Learning of malicious behavior vocabulary and threat detection through behavior matching

Inverse imbalance subspace searching techniques are used to detect potential malware among samples of network communication data. A large number of samples of network communication data, such as proxy log data and/or network flows, are received and analyzed by a malware detection system. A number of the samples are associated with known malware, while other unlabeled samples are either benign or may be associated with unknown malware. An inverse imbalance subspace search may be performed, in which the sample sets are divided into subsets based on random feature thresholds, and each subset is evaluated based on the ratio of known malware samples to unlabeled samples. Unlabeled samples within subsets having high malware sample ratios may be identified, aggregated, and processed as potential malware.

Malware detection using inverse imbalance subspace searching

Techniques are described for detecting attacks that employ a display name in an email to impersonate an email sender. A computing infrastructure hosting an email security platform may determine a similarity between the display name and an email address from which the email was received. The email security platform may determine the similarity by comparing a string associated with the display name and a string associated with the sender address. The email security platform may generate a similarity value based on a result of the display name being compared with the sender address. The email security platform may determine that the email includes the display name impersonating a name of the sender, based on the similarity value meeting or exceeding a threshold value indicative of impersonation. The email security platform may delete or quarantine the email from an inbox associated with a user account.

Email security based on display name and address

In one embodiment, a device distributes sets of training records from a training dataset for a random forest-based classifier among a plurality of workers of a computing cluster. Each worker determines whether it can perform a node split operation locally on the random forest by comparing a number of training records at the worker to a predefined threshold. The device determines, for each of the split operations, a data size and entropy measure of the training records to be used for the split operation. The device applies a machine learning-based predictor to the determined data size and entropy measure of the training records to be used for the split operation, to predict its completion time. The device coordinates the workers of the computing cluster to perform the node split operations in parallel such that the node split operations in a given batch are grouped based on their predicted completion times.

Distributed random forest training with a predictor trained to balance tasks

In one embodiment, a traffic analysis service obtains telemetry data regarding encrypted traffic associated with a particular device in the network, wherein the telemetry data comprises Transport Layer Security (TLS) features of the traffic. The service determines, based on the TLS features from the obtained telemetry data, a set of one or more TLS fingerprints for the traffic associated with the particular device. The service calculates a measure of similarity between the set of one or more TLS fingerprints for the traffic associated with the particular device and a set of one or more TLS fingerprints of traffic associated with a second device. The service determines, based on the measure of similarity, that the particular device and the second device were operated by the same user.

Device detection in network telemetry with TLS fingerprinting

In one embodiment, a device trains a machine learning-based malware classifier using a first randomly selected subset of samples from a training dataset. The classifier comprises a random decision forest. The device identifies, using at least a portion of the training dataset as input to the malware classifier, a set of misclassified samples from the training dataset that the malware classifier misclassifies. The device retrains the malware classifier using a second randomly selected subset of samples from the training dataset and the identified set of misclassified samples. The device adjusts prediction labels of individual leaves of the random decision forest of the retrained malware classifier based in part on decision changes in the forest that result from assessing the entire training dataset with the classifier. The device sends the malware classifier with the adjusted prediction labels for deployment into a network.

Scalable training of random forests for high precise malware detection

In one embodiment, a computing device provides a feature vector as input to a random decision forest comprising a plurality of decision trees trained using a training dataset, each decision tree being configured to output a classification label prediction for the input feature vector. For each of the decision trees, the computing device determines a conditional probability of the decision tree based on a true classification label and the classification label prediction from the decision tree for the input feature vector. The computing device generates weightings for the classification label predictions from the decision trees based on the determined conditional probabilities. The computing device applies a final classification label to the feature vector based on the weightings for the classification label predictions from the decision trees.

Bayesian tree aggregation in decision forests to increase detection of rare malware

In one embodiment, a computing device trains a multi-class classifier (having a plurality of classes) on a training dataset, and evaluates the multi-class classifier on a testing dataset to determine a performance of each of the plurality of classes. The plurality of classes may then be partitioned into either learnable or unlearnable based on whether the performance each particular class surpasses a particular threshold, and then a predicting classifier can be trained on the training dataset, where data of the training dataset is labelled as either learnable or unlearnable based on the particular class to which the data corresponds. Accordingly, the computing device may then use the predicting classifier on a new class to predict whether samples associated with the new class are learnable or unlearnable, and may retrain the multi-class classifier with the samples associated with the new class in response to predicting that the samples are learnable.

Growing community of inventors

Rakovnik, Czechia

Jan Brabec