International Business Machines Corporation

Other

Benjamin J Edwards

Ian M Molloy

Taesung Lee

Suresh N Chari

Bryant R Chen

Youngja Park

Heiko Hary Ludwig

Jialong Zhang

Dong Su

Wilka Carvalho

Heqing Huang

Ebube Chuba

One or more computer processors determine a tolerance value, and a norm value associated with an untrusted model and an adversarial training method. The one or more computer processors generate a plurality of interpolated adversarial images ranging between a pair of images utilizing the adversarial training method, wherein each image in the pair of images is from a different class. The one or more computer processors detect a backdoor associated with the untrusted model utilizing the generated plurality of interpolated adversarial images. The one or more computer processors harden the untrusted model by training the untrusted model with the generated plurality of interpolated adversarial images.

Adversarial interpolation backdoor detection

Mechanisms are provided to implement a hardened ensemble artificial intelligence (AI) model generator. The hardened ensemble AI model generator co-trains at least two AI models. The hardened ensemble AI model generator modifies, based on a comparison of the at least two AI models, a loss surface of one or more of the at least two AI models to prevent an adversarial attack on one AI model, in the at least two AI models, transferring to another AI model in the at least two AI models, to thereby generate one or more modified AI models. At least one of the one or more modified AI models then processes an input to generate an output result.

Training diverse and robust ensembles of artificial intelligence computer models

Mechanisms are provided to implement a hardened neural network framework. A data processing system is configured to implement a hardened neural network engine that operates on a neural network to harden the neural network against evasion attacks and generates a hardened neural network. The hardened neural network engine generates a reference training data set based on an original training data set. The neural network processes the original training data set and the reference training data set to generate first and second output data sets. The hardened neural network engine calculates a modified loss function of the neural network, where the modified loss function is a combination of an original loss function associated with the neural network and a function of the first and second output data sets. The hardened neural network engine trains the neural network based on the modified loss function to generate the hardened neural network.

Deep neural network hardening framework

One embodiment provides a method comprising receiving a training set comprising a plurality of data points, where a neural network is trained as a classifier based on the training set. The method further comprises, for each data point of the training set, classifying the data point with one of a plurality of classification labels using the trained neural network, and recording neuronal activations of a portion of the trained neural network in response to the data point. The method further comprises, for each classification label that a portion of the training set has been classified with, clustering a portion of all recorded neuronal activations that are in response to the portion of the training set, and detecting one or more poisonous data points in the portion of the training set based on the clustering.

Detecting poisoning attacks on neural networks by activation clustering

Mechanisms are provided for evaluating a trained machine learning model to determine whether the machine learning model has a backdoor trigger. The mechanisms process a test dataset to generate output classifications for the test dataset, and generate, for the test dataset, gradient data indicating a degree of change of elements within the test dataset based on the output generated by processing the test dataset. The mechanisms analyze the gradient data to identify a pattern of elements within the test dataset indicative of a backdoor trigger. The mechanisms generate, in response to the analysis identifying the pattern of elements indicative of a backdoor trigger, an output indicating the existence of the backdoor trigger in the trained machine learning model.

Using gradients to detect backdoors in neural networks

Jointly discovering user roles and data clusters using both access and side information by performing the following operation: (i) representing a set of users as respective vectors in a user feature space; representing data as respective vectors in a data feature space; (ii) providing a user-data access matrix, in which each row represents a user's access over the data; and (iii) co-clustering the users and data using the user-data matrix to produce a set of co-clusters.

Jointly discovering user roles and data clusters using both access and side information

An approach is provided that receives a set of user information pertaining to a user. The received set of information is encoded into a neural network and the neural network is trained using the encoded user information. As an output of the trained neural network, passwords corresponding to the user are generated.

Deep learning for targeted password generation with cognitive user information understanding

Mechanisms are provided to implement an adversarial network framework. Using an adversarial training technique, an image obfuscation engine operating as a generator in the adversarial network framework is trained to determine a privacy protection layer to be applied by the image obfuscation engine to input image data. The image obfuscation engine applies the determined privacy protection layer to an input image captured by an image capture device to generate obfuscated image data. The obfuscated image data is transmitted to a remotely located image recognition service, via at least one data network, for performance of image recognition operations.

Adversarial learning of privacy protection layers for image recognition services

Detecting anomalous user behavior is provided. User activity is logged for a set of users. The user activity is divided into distinct time intervals. For each distinct time interval, logged user activity is converted to a numerical representation of each user's activities for that distinct time interval. A clustering process is used on the numerical representations of user activities to determine which users have similar activity patterns in each distinct time interval. A plurality of peer groups of users is generated based on determining the similar activity patterns in each distinct time interval. Anomalous user behavior is detected based on a user activity change in a respective peer group of users within a distinct time interval.

Clustering for detection of anomalous behavior and insider threat

A smoker's booth, for use within an enclosed area, including sidewalls having a door providing access to a single user compartment, wherein the sidewalls define upper and lower open ends. The lower end is closed by a base panel supporting the sidewalls, while the upper end is closed by a modular cap assembly, adapted for treating tobacco smoke-laden air drawn upwardly from the closed compartment. The modular cap assembly, which includes a closed space divided into a plurality of chambers, provides at least one plenum chamber and an electrical equipment chamber. An air treating unit is located in the plenum chamber, intermediate intake and return air vent openings communicating with the compartment. A blower draws tobacco smoke-laden air upwardly from the compartment, through the intake vent opening and plenum chamber, and returns only treated air to the compartment, via its return vent. As the compartment air is self contained, the booth obviates exchange of ambient air with tobacco smoke-laden air during its operation. A second embodiment provides an adjustable damper in the plenum chamber, allowing all or a selected portion of the treated air to be exhausted to the ambient, while an ambient air intake louver, in the booth sidewall, draws an amount of ambient 'make-up' air into the compartment equal to the amount of exhausted treated air.

Growing community of inventors

Ossining, NY, United States of America

Benjamin J Edwards