Performing privacy-preserving multi-party analytics on horizontally partitioned local data

1. A method comprising:

computing, by a computing device at a party among a plurality of parties, a sum of local data owned by the party, wherein the local data is horizontally partitioned into a plurality of data segments, and wherein each data segment comprises a non-overlapping subset of data entries owned by a particular party;

computing, by the computing device, a local gradient based on the horizontally partitioned local data;

initializing, by the computing device, each data segment by:

dividing each data segment into at least two data shards;

recursively encrypting each data shard using a plurality of public keys corresponding to the plurality of parties and the mediator; and

sharing each encrypted data shard with the mediator;

receiving, by the computing device, an aggregated local gradient from a mediator, wherein the mediator computes the aggregated local gradients based on a plurality of data entries in the each initialized data segment owned by the plurality of parties;

anonymizing, by the computing device, the aggregated local gradient received from the mediator;

receiving, by the computing device from the mediator, a global gradient based on the anonymized aggregated local gradients;

learning, by the computing device, a global analytic model based on the global gradient; and

performing, by the computing device, privacy-preserving multi-party analytics on the horizontally partitioned local data based on the learned global analytic model.

2. The method of claim 1 , wherein computing the local gradient based on the horizontally partitioned local data uses a gradient descent function to search for a solution by iteratively performing a plurality of operations in a direction of negative gradient.

3. The method of claim 2 , wherein the plurality of operations comprise initializing local model parameters, and iteratively updating the local model parameters for each dimension based on a current gradient value until a termination criteria is satisfied, wherein the local model parameters are adapted to the global analytic model upon the termination criteria being satisfied.

4. The method of claim 1 , wherein the cost function comprises one of: a linear regression cost function, a logistic regression cost function, a ridge regression cost function, a support vector machine (SVM) function, and a least absolute shrinkage and selection operator (LASSO) function.

5. The method of claim 1 , wherein the plurality of parties participate in data sharing under a semi-honest adversary model in which each party follows an agreed-upon protocol while attempting to acquire information about private data during computation.

6. The method of claim 1 , wherein anonymizing the aggregated local data further comprises:

decrypting, by the computing device, a collection of data shards received from the mediator using the private key corresponding to the first party;

shuffling, by the computing device, the collection of data shards that are decrypted using the private key corresponding to the first party; and

transmitting, by the computing device, the shuffled collection of data shards to one of the plurality of parties and the mediator.

7. The method of claim 6 , wherein the collection of data shards is further decrypted using other private keys corresponding to other parties in the plurality of parties following a predetermined priority order.

8. A system comprising:

a plurality of computing devices associated with a plurality of parties participating in privacy-preserving data sharing under a semi-honest adversary model, and

a mediator coupled to the plurality of parties;

wherein each of the plurality of computing devices comprises a hardware processor to:

transform a cost function of a data analytics task to a gradient descent function;

compute a local gradient of local data owned by a particular party based on the cost function, wherein the local data is horizontally partitioned into a plurality of data segments, and wherein each data segment comprises a non-overlapping subset of data entries owned by the particular party;

initialize each data segment by:

dividing each data segment into at least two data shards;

recursively encrypting each data shard using a plurality of public keys corresponding to the plurality of parties and the mediator; and

sharing each encrypted data shard with the mediator;

receive an aggregated local gradient from a mediator, wherein the mediator computes the aggregated local gradients based on a plurality of data entries in the each initialized data segment;

anonymize the aggregated local gradient received from the mediator;

receive from a mediator a global gradient computed based on the anonymized aggregated local gradients;

learn a global analytic model based on the global gradient; and

wherein the mediator computes a global gradient used in the global analytic model to perform privacy-preserving multi-party analytics on the horizontally partitioned local data.

9. The system of claim 8 , wherein the gradient descent function searches for a solution by iteratively performing a plurality of operations in a direction of negative gradient.

10. The system of claim 9 , wherein the plurality of operations comprise initializing the local model parameters, and iteratively updating the local model parameters for each dimension based on a current gradient value until a termination criteria is satisfied, wherein the local model parameters are adapted to the global analytic model upon the termination criteria being satisfied.

11. The system of claim 8 , wherein the plurality of parties participate in data sharing under a semi-honest adversary model in which each party follows an agreed-upon protocol while attempting to acquire information about private data during computation.

12. The system of claim 8 , wherein the hardware processor in each of the plurality of computing devices is further to:

decrypt a collection of data shards received from the mediator using the private key corresponding to a first party;

shuffle the collection of data shards that are decrypted using the private key corresponding to the first party; and

transmit the shuffled collection of data shards to one of the plurality of parties and the mediator.

13. The system of claim 12 , wherein the collection of data shards is further decrypted using other private keys corresponding to other parties in the plurality of parties following a predetermined priority order.

14. The system of claim 8 , wherein the cost function comprises one of: a linear regression cost function, a logistic regression cost function, a ridge regression cost function, a support vector machine (SVM) function, and a least absolute shrinkage and selection operator (LASSO) function.

15. A non-transitory machine-readable storage medium encoded with instructions executable by at least one processor of a computing device associated with a mediator in a multi-party environment, the machine-readable storage medium comprising instructions to:

receive, from a plurality of parties, local gradients computed based on local data owned by the party, wherein the local data is horizontally partitioned into a plurality of data segments, and wherein each data segment comprises a non-overlapping subset of data entries;

divide each data segment into at least two data shards;

recursively encrypt each data shard using a plurality of public keys corresponding to the plurality of parties and the mediator;

share each encrypted data shard with the mediator;

transform a cost function of a data analytics task to a gradient descent function based on the received local gradients;

aggregate the local gradients received from the plurality of parties into aggregated local gradients;

compute a global gradient based on the aggregated local gradients;

share the aggregated local gradients and the global gradient with the plurality of parties wherein each party learns a global analytic model, wherein the aggregated local gradients comprise a plurality of cost function values for data entries owned by the plurality of parties; and

perform privacy-preserving multi-party analytics on horizontally partitioned local data based on the global analytic model.

16. The non-transitory machine-readable storage medium of claim 15 , wherein the gradient descent function searches for a solution by iteratively performing a plurality of operations in a direction of negative gradient.

17. The non-transitory machine-readable storage medium of claim 16 , wherein the plurality of operations comprise initializing the local model parameters, and updating the local model parameters for each dimension based on a current gradient value until a termination criteria is satisfied.

18. The non-transitory machine-readable storage medium of claim 15 , wherein the cost function comprises one of a linear regression cost function, a logistic regression cost function, a ridge regression cost function, a support vector machine (SVM) function, and a least absolute shrinkage and selection operator (LASSO) function.