International Business Machines Corporation

Marc Stoecklin

Andreas Kind

Ting Wang

Douglas Lee Schales

Reiner Sailer

Paul T Hurley

Mihai Christodorescu

Ian M Molloy

Jiyong Jang

Jialong Zhang

Zhongshu Gu

Patrick Droz

Olaf W Zimmermann

Xenofontas Dimitropoulos

A method, apparatus and computer program product to protect a deep neural network (DNN) having a plurality of layers including one or more intermediate layers. In this approach, a training data set is received. During training of the DNN using the received training data set, a representation of activations associated with an intermediate layer is recorded. For at least one or more of the representations, a separate classifier (model) is trained. The classifiers, collectively, are used to train an outlier detection model. Following training, the outliner detection model is used to detect an adversarial input on the deep neural network. The outlier detection model generates a prediction, and an indicator whether a given input is the adversarial input. According to a further aspect, an action is taken to protect a deployed system associated with the DNN in response to detection of the adversary input.

Detecting adversary attacks on a deep neural network (DNN)

A method for identifying an unknown user according to a plurality of facets of user activity in a plurality of contexts includes receiving a plurality of priors for the facets with respect to the contexts, receiving a plurality of footprints of known users, aggregating the footprints of the users to determine an ensemble prior, receiving a plurality of network traces relevant to an unknown user in a computer environment, matching the network traces against each of the footprints to determine a plurality of matches, aggregating the matches using the ensemble prior according to the facets and the contexts, and outputting a probable user identity for the unknown user.

User identification using multifaceted footprints

A method for probabilistic lossy counting includes: for each element in a current window, determining whether an entry corresponding to a current element is present in a table; in the event an entry corresponding to the current element is present in the table, incrementing a frequency counter associated with the current element; otherwise, inserting an entry into a table, wherein inserting an entry comprises: calculating a probabilistic error bound &#x394; based on an index i of the current window; and inserting the probabilistic error bound &#x394; and a frequency counter into an entry corresponding to the current element in the table; and at the end of the current window, removing all elements from the table wherein the sum of the frequency counter and probabilistic error bound &#x394; associated with the element is less than or equal to the index of the current window.

Probabilistic lossy counting

A system and method for monitoring packetized traffic flow in a network and enabling approximation of the rate information of a network flow. The method for monitoring network traffic flow includes receiving, at a network packet flow collector device, packetized traffic flow signals to be monitored; sampling said received packetized traffic flow signals in time to form an approximation of the packet flow rate in time; generating packet flow activity data comprising data representing the sampled traffic flow signals sampled in time; communicating the packet flow activity data to a network packet flow analyzer device and processing the flow activity data to form signals representing an approximate version of the network traffic flow in the network, the analyzer processing the traffic flow signals for reconstructing the rate of the netflow as a function of time. The flow analyzer then generates a compressed version of the network traffic flow signals in the network, the compressed network traffic flow signals comprising relevant approximation of the packet flow rate in time.

System and method for network flow traffic rate encoding

Systems, methods, and computer program products for extracting port-level information of Web services with flow-based network monitoring. Exemplary embodiments include a method for extracting port-level information of Web services with flow-based network monitoring, the method including identifying a registry machine, coupling the registry machine to a traffic meter and flow monitor dynamically configuring the traffic meter, including exporting a first n bytes of a traffic payload exporting a sub-second traffic flow start and end times, extracting service provider information from traffic flow exports, including analyzing the exported n bytes of the traffic payload to extract port-level information at the flow monitor, extracting a value of an access point element, mapping a logical service provider address to a physical address and inserting the service into a Web Service Provider Registry within the flow meter, thereby populating the Web Service Provider Registry.

Systems, methods and computer program products for extracting port-level information of web services with flow-based network monitoring

Growing community of inventors

Basel, Switzerland

Marc Stoecklin