International Business Machines Corporation

Ehud Aharoni

Allon Adir

Omri Soceanu

Lev Greenberg

Oded Margalit

Noam Slonim

Hayim Shaul

Nir Drucker

Gilad Ezov

Dan Gutfreund

Ramy Masalha

Ruty Rinott

Tamar Lavee

Eyal Kushnir

Ran Levy

Daniel Hershcovich

Ariel Farkash

Tamer Salman

Yonatan Bilu

Dov Murik

Robert J Baseman

Oded Sofer

Indrajit Bhattacharya

Raminderpal Singh

Adam D Ticknor

Shmuel Regev

Amir Ronen

Hong Lin

Boris Rozenberg

Fateh A Tipu

Ronen Levy

Hani Neuvirth-Telem

Ron Peleg

Michal Rosen-Zvi

Ramona Kei

Kevin MacKey

Timothy M McCormack

Michael Mirkin

Rosa Miroshnikov

Roza Miroshnikov

Asaf Polakovski

Levente Klein

Roman Vaculin

Heiko Hary Ludwig

David Carmel

Nathalie Baracaldo Angel

James Thomas Rayfield

Amrita Saha

Onn Menahem Shehory

Yehuda Naveh

Vikas Raykar

Kanthi K Sarpatwar

Akram Bitar

Roy N Bar-Haim

Upendra D Chitnis

Evgeny Shindin

Chen Yanover

Guy Moshkowich

Yaara Goldschmidt

Runhua Xu

Saharon Rosset

Lena Dankin

Hani Neuvirth

Ya'ara Goldschmidt

Francesco Dinuzzo

Godbole Shantanu

A technique for privacy-preserving homomorphic inferencing using a neural network having an activation function, such as a non-linear high-degree polynomial. The network is trained to learn input features of an input feature vector together with their associated inverses. During inferencing, an encrypted data point is received at the network. The data point comprises an input feature vector that has been extended with a set of one or more additional feature values, the set of one or more additional feature values having been generated by applying a normalized inverse function to respective one or more features in the feature vector. Homomorphic inferencing is performed on the encrypted data point using the machine learning model to generate an encrypted result, which is then returned. By applying the normalized inverse function, the high-degree polynomial can use any value of an input feature during inferencing, whether the value is within or outside of a particular input range.

Non-linear approximation robust to input range of homomorphic encryption analytics

An embodiment appends, into a concatenated table, a second plurality of records in a second table to a first plurality of records in a first table. An embodiment sorts, according to each identification value in the concatenated table, the concatenated table. An embodiment generates, using an equality mask derived from each identification value in the sorted table, an intersection table, the intersection table comprising a record in the first plurality of records with a first identifier value matching a second identifier value in a record in the second plurality of records. An embodiment generates, using a not-in-intersection mask derived from the equality mask, a not-in-intersection table. An embodiment adds contents of the intersection table and contents of the not-in-intersection table together, resulting in a join table.

Database table joining under fully homomorphic encryption

Mechanisms are provided for performing a tournament-league comparison process of a computer function. A request is received to execute the computer function on a vector data structure, where a result of the computer function is provided by executing the tournament-league comparison process. The vector data structure comprises a plurality of values where each value corresponds to a vector slot. At least one iteration of a tournament comparison operation is executed to generate a first intermediate ciphertext and indicator matrix, where the first intermediate ciphertext comprises fewer vector slots than the at least one input vector data structure. A plurality of iterations of a league comparison operation are executed based on the first intermediate ciphertext and one or more second intermediate ciphertexts generated at each iteration of the league comparison operation. A final iteration of the league comparison operation is executed that outputs a final result of the tournament-league comparison process.

Tournament-league mechanisms for comparison based computer operations

An example system includes a processor to mask a ciphertext using four random elements to generate masked ciphertexts. The processor can send the masked ciphertexts to a server device. The processor can receive masked plaintexts from the server device. The processor can unmask the masked plaintexts using the four random elements to generate unmasked plaintexts.

Masking ciphertexts using random elements for secure decryption

An example system includes a processor to receive a ciphertext including an encrypted packed plaintext. The processor can, in response to detecting that a reshape is to be executed for a data structure including the packed ciphertext, identify a target shape for the data structure and execute the reshape based on the identified target shape. The reshape is executed using multiplication by an identity element. The processor can then execute a homomorphic computation using the reshaped data structure.

Reshaping data structures for homomorphic computations using multiplication by identity elements

An example system includes a processor to compute a tensor of indicators indicating a presence of partial sums in an encrypted vector of indicators. The processor can also securely reorder an encrypted array based on the computed tensor of indicators to generate a reordered encrypted array.

Secure reordering using tensor of indicators

A computer-implemented method for generating hash values to determine string similarity is disclosed. The computer-implemented method includes converting a first text string of a first data set into a first set of shingles. The computer-implemented method further includes determining a weight associated with each shingle in the first set of shingles based, at least in part, on a particular record field associated with a shingle. The computer-implemented method further includes generating, based on a hash function, a hash value for each shingle in the first set of shingles. The computer-implemented method further includes reducing the hash value generated for each shingle in the first set of shingles, based, at least in part on the weight associated with the shingle.

String similarity based weighted min-hashing

Mechanisms are provided for performing a tournament selection process of a computer function. A request is received to execute the computer function on an input vector data structure, where a result of the computer function is provided by executing the tournament selection process. The input vector data structure is received, comprising a plurality of values where each value corresponds to a vector slot. An index vector data structure is received that comprises indices of the vector slots of the input vector. Iteration(s) of the tournament selection process are executed to identify a value in the input vector satisfying a criterion of the computer function. An operation is performed on the index vector data structure to generate an indicator vector data structure that uniquely identifies a slot in the input vector data structure that is a result of the computer function being executed on the input vector data structure.

Tournament type selection operations on encrypted data

A method and system for evaluating and selecting an optimal packing solution (or solutions) for data that is run through a fully homomorphic encryption (FHE) simulation. In some instances, a user selected model architecture is provided in order to start simulating multiple potential configurations. Additionally, the cost of each simulated configuration is taken into account when determining an optimal packing solution.

Optimal profile selection for FHE based analytical models

A computer-implemented method comprising: generating, from a key-seed associated with a user, a set of homomorphic encryption (HE) keys associated with an HE scheme; receiving, from a key management system (KMS) associated with said HE scheme, an encrypted version of said key-seed; storing said encrypted version of said key-seed, and said set of HE keys, in an untrusted storage location; and at a decryption stage, decrypting an encrypted computation result generated using said HE scheme, by: (i) recalling, from said untrusted storage location, said encrypted version of said key-seed, (ii) providing said encrypted version of said key-seed to said KMS, to obtain a decrypted version of said key-seed s associated with said user, (iii) generating, from said received decrypted version of said key-seed, a secret HE key associated with said HE scheme, and (iv) using said secret HE key to decrypt said encrypted computation result.

Homomorphic encryption key management

A method, apparatus and computer program product for privacy-preserving homomorphic inferencing. In response to receipt of encrypted data, a ciphertext of real numbers is generated. Each real number has an associated sign that is desired to be maintained. A mask is then identified, preferably via an iterative algorithm that works on a trial and error basis to locate an appropriate solution. The mask comprises set of values randomly distributed over a given positive range and that remain positive after encoding under a fixed-point arithmetic and with a low scale value. Under homomorphic encryption, the ciphertext is then multiplied by the mask to generate a result comprising values corresponding to the real numbers in the ciphertext and that maintain their associated signs. The result is provided as a response to the encrypted data.

Efficient random masking of values while maintaining their sign under fully homomorphic encryption (FHE)

Mechanisms are provided for fully homomorphic encryption enabled graph embedding. An encrypted graph data structure, having encrypted entities and predicates, is received and, for each encrypted entity, a corresponding set of entity ciphertexts is generated based on an initial embedding of entity features. For each encrypted predicate, a corresponding predicate ciphertext is generated based on an initial embedding of predicate features. A machine learning process is iteratively executed, on the sets of entity ciphertexts and the predicate ciphertexts, to update embeddings of the entity features of the encrypted entities and update embeddings of predicate features of the encrypted predicates, to generate a computer model for embedding entities and predicates. A final embedding is output based on the updated embeddings of the entity features and predicate features of the computer model.

Graph based embedding under fully homomorphic encryption

A method, a neural network, and a computer program product are provided that optimize training of neural networks using homomorphic encrypted elements and dropout algorithms for regularization. The method includes receiving, via an input to the neural network, a training dataset containing samples that are encrypted using homomorphic encryption. The method also includes determining a packing formation and selecting a dropout technique during training of the neural network based on the packing technique. The method further includes starting with a first packing formation from the training dataset, inputting the first packing formation in an iterative or recursive manner into the neural network using the selected dropout technique, with a next packing formation from the training dataset acting as an initial input that is applied to the neural network for a next iteration, until a stopping metric is produced by the neural network.

Neural network training with homomorphic encryption

Privacy-preserving homomorphic inferencing utilizes batch processing on encrypted data records. Each data record has a private data portion of interest against which the inferencing is carried out. Batch processing is enabled with respect to a set of encrypted data records by techniques that ensure that each encrypted data record has its associated private data portion in a unique location relative to the other data records. The set of encrypted data records are then summed to generate a single encrypted data record against which the inferencing is done. In a first embodiment, the private data portions of interest are selectively and uniquely positioned at runtime (when the inferencing is being applied). In a second embodiment, the private data portions of interest are initially positioned with the data-at-rest, preferably in an off-line process; thereafter, at runtime individual encrypted data records are processed as necessary to adjust the private data portions to unique positions prior to batching.

Efficiently batching pre-encrypted data for homomorphic inference

A method comprising: receiving an input tensor having a shape defined by [n, . . . , n], where k is equal to a number of dimensions that characterize the input tensor; receiving tile tensor metadata comprising: a tile tensor shape defined by [t, . . . , t], and information indicative of an interleaving stride to be applied with respect to each dimension of the tile tensor; constructing an output tensor comprising a plurality of the tile tensors, by applying a packing algorithm which maps each element of the input tensor to at least one slot location of one of the plurality of tile tensors, based on the tile tensor shape and the interleaving stride, wherein the interleaving stride results in non-contiguous mapping of the elements of the input tensor, such that each of the tile tensors includes a subset of the elements of the input tensor which are spaced within the input tensor according to the interleaving stride.

Efficient convolution in an environment that enforces tiles

Growing community of inventors

Kfar Saba, Israel

Ehud Aharoni