International Business Machines Corporation

Ron Shmelkin

Abigail Goldsteen

Ariel Farkash

Gilad Ezov

Micha Gideon Moffie

Muhammad Barham

Omri Soceanu

Michael Vinov

An embodiment analyzes a predictive model and its input data for the predictive model using an explainability algorithm resulting in a feature importance value of a feature. The embodiment analyzes feature values of the feature using a generalization function resulting in a set of candidate feature values. The embodiment determines an alternative feature based on the set of candidate feature values, wherein the alternative feature is a generalization of the feature. The embodiment compares an accuracy of the predictive model to a threshold performance value and, responsive to the accuracy being above the threshold performance value, maps feature values in the input data that are in the set of candidate feature values to a generalized representative value in the generalized domain.

Data minimization using global model explainability

A method, apparatus and a product for data generalization for predictive models. The method comprising: obtaining a training dataset that comprises a plurality of training instances and predicted labels thereof, wherein each training instance is a valuation of a set of features, wherein the set of features comprises a feature having a domain, wherein the predicted label of each training instance is a label predicted thereto by a predictive model; training an auxiliary model using the training dataset; based on the auxiliary model, determining an alternative set of features that is a generalization of the set of features, wherein the alternative set of features comprises a generalized feature having a generalized domain, wherein each value in the generalized domain corresponds to one or more values in the domain; obtaining a generalized instance having a valuation of the alternative set of features; and determining a label for the generalized instance.

Data generalization for predictive models

A method, computer system, and a computer program product for testing a data removal from a trained machine learning model trained with a training data set are provided. A new machine learning model is trained by using an altered data set that includes training data from the training data set. The altered data set is without removal data. A first forgetting mechanism is applied to the trained machine learning model to form a first revised machine learning model. The applying includes removing the removal data from the trained machine learning model. A first membership leakage quantification on the first revised machine learning model is performed to quantify a first membership leakage of the removal data and that uses the new machine learning model for comparison. A first leakage score is determined from the first membership leakage quantification to test the forgetting mechanism.

Membership leakage quantification to verify data removal

An example system includes a processor to receive training data and predictions on the training data of a trained machine learning model to be anonymized. The processor is to generate generalized data from training data based on the predictions of the trained machine learning model on the training data. The processor is to train an anonymized machine learning model using the generalized data.

Training anonymized machine learning models via generalized data generated using received trained machine learning models

A method, apparatus and a product for data generalization for predictive models. The method comprising: based on a labeled dataset, determining a plurality of buckets, each of which has an associated label; determining a plurality of clusters, grouping similar instances in the same bucket; based on the plurality of clusters, determining an alternative set of features comprising a set of generalized features, wherein each generalized feature corresponds to a cluster of the plurality of clusters, wherein a generalized feature that corresponds to a cluster is indicative of the instance being mapped to the corresponding cluster; obtaining a second instance; determining a generalized second instance that comprises a valuation of the alternative set of features for the second instance; and based on the generalized second instance, determining a label for the second instance.

Embodiments of the present systems and methods may provide encrypted biometric information that can be stored and used for authentication with undegraded recognition performance. For example, in an embodiment, a method may comprise storing a plurality of encrypted trained weights of a neural network classifier, wherein the weights have been trained using biometric information representing at least one biometric feature of a person, receiving encrypted biometric information obtained by sampling at least one biometric feature of the person and encrypting the sampled biometric feature, obtaining an match-score using the encrypted trained neural network classifier, the match-score indicating a probability that the received encrypted biometric information matches the stored encrypted biometric information, and authenticating the person when the probability that received encrypted biometric information matches the stored encrypted biometric information exceeds a threshold.

Method and system for privacy preserving biometric authentication

A method of limiting data usage for certified purposes by using functional encryption, comprising: receiving from a software publisher an application code and declared privacy information, the declared privacy information specifies at least one declared usage for at least one data type; analyzing the application's usage of data collected by the application, to identify an actual usage of the at least one data type by a function; identifying when the actual usage is compliant with the at least one declared usage according to the analysis; in response to the identification, creating a pair of a public key and a master private key; creating a function private key for the function using the master private key; and sending the function private key to the software publisher to be used for operating the function on data which is encrypted using the public key.

Growing community of inventors

Haifa, Israel

Ron Shmelkin