Verifying remote execution of machine learning inference under homomorphic encryption using permutations
A technique to remotely identify potential compromise of a service provider that performs homomorphic inferencing on a model. For a set of real data samples on which the inferencing is to take place, at least first and second permutations of a set of trigger samples are generated. Every set of samples (both trigger and real samples) are then sent for homomorphic inferencing on the model at least twice, and in a secret permutated way. To improve performance, a permutation is packaged with the real data samples prior to encryption using a general purpose data structure, a tile tensor, that allows users to store multi-dimensional arrays (tensors) of arbitrary shapes and sizes. In response to receiving one or more results from the HE-based model inferencing, a determination is made whether the service provider is compromised. Upon a determination that the service provider is compromised, a given mitigation action is taken.
1 . A method for verifying proof of work of a service provider that performs homomorphic inferencing on a model, comprising:
generating a set of trigger samples;
packing the set of trigger samples together with each of first and second sets of real data samples, wherein the set of trigger samples are packed with the first set of real data samples in first encrypted data according to a first permutation, and wherein the set of trigger samples are packed with the second set of real data samples in second encrypted data according to a second permutation that differs from the first permutation;
forwarding the first and second encrypted data to the service provider for homomorphic encryption-based inferencing on the model;
receiving from the service provider one or more results;
based on the one or more results, determining whether the service provider is compromised; and
upon a determination that the service provider is compromised, taking a given mitigation action.
2 . The method as described in claim 1 wherein an ordering of the trigger samples in each of the first and second permutations is secret except to a verifier that determines whether the service provider is compromised.
3 . The method described in claim 2 wherein determining whether the service provider is compromised includes the verifier validating that at least one label predicated by the model for a sample in the set of trigger samples matches a label returned in the one or more results.
4 . The method as described in claim 2 wherein the verifier is one of: the data owner, an owner of the model, and a combination of the data owner and the model owner.
5 . The method as described in claim 1 wherein a sample in the set of trigger samples is one of: noise, a seed applied to a random number generator, and a data sample associated with at least one of the first second sets of real data samples and that has a given relationship to a classification boundary.
6 . The method as described in claim 1 wherein the first and second permutations are packed with the first and second set of real data samples as a tile tensor.
7 . An apparatus, comprising:
at least one processor;
computer memory holding computer program instructions executed by the at least one processor to verify proof of work of a service provider that performs homomorphic inferencing on a model, the computer program instructions comprising program code configured to:
generate a set of trigger samples;
pack the set of trigger samples together with each of first and second sets of real data samples, wherein the set of trigger samples are packed with the first set of real data samples in first encrypted data according to a first permutation, and wherein the set of trigger samples are packed with the second set of real data samples in second encrypted data according to a second permutation that differs from the first permutation;
forward the first and second encrypted data to the service provider for homomorphic encryption-based inferencing on the model;
receive from the service provider one or more results;
based on the one or more results, determine whether the service provider is compromised; and
upon a determination that the service provider is compromised, take a given mitigation action.
8 . The apparatus as described in claim 7 wherein an ordering of the trigger samples in each of the first and second permutations is secret except to a verifier that determines whether the service provider is compromised.
9 . The apparatus as described in claim 8 wherein the program code configured to determine whether the service provider is compromised includes program code to validate that at least one label predicated by the model for a sample in the set of trigger samples matches a label returned in the one or more results.
10 . The apparatus as described in claim 8 wherein the processing system is associated with a verifier that is one of: the data owner, an owner of the model, and a combination of the data owner and the model owner.
11 . The apparatus as described in claim 7 wherein a sample in the set of trigger samples is one of: noise, a seed applied to a random number generator, and a data sample associated with at least one of the first second sets of real data samples and that has a given relationship to a classification boundary.
12 . The apparatus as described in claim 7 wherein the first and second permutations are packed with the first and second set of real data samples as a tile tensor.
13 . A computer program product in a non-transitory computer readable medium, the computer program product holding computer program instructions executed by at least one processor in a processing system to verify proof of work of a service provider that performs homomorphic inferencing on a model, the computer program instructions comprising program code configured to:
generate a set of trigger samples;
pack the set of trigger samples together with each of first and second sets of real data samples, wherein the set of trigger samples are packed with the first set of real data samples in first encrypted data according to a first permutation, and wherein the set of trigger samples are packed with the second set of real data samples in second encrypted data according to a second permutation that differs from the first permutation;
forward the first and second encrypted data to the service provider for homomorphic encryption-based inferencing on the model;
receive from the service provider one or more results;
based on the one or more results, determine whether the service provider is compromised; and
upon a determination that the service provider is compromised, take a given mitigation action.
14 . The computer program product as described in claim 13 wherein an ordering of the trigger samples in each of the first and second permutations is secret except to a verifier that determines whether the service provider is compromised.
15 . The computer program product as described in claim 14 wherein the program code configured to determine whether the service provider is compromised includes program code to validate that at least one label predicated by the model for a sample in the set of trigger samples matches a label returned in the one or more results.
16 . The computer program product as described in claim 14 wherein the processing system is associated with a verifier that is one of: the data owner, an owner of the model, and a combination of the data owner and the model owner.
17 . The computer program product as described in claim 13 wherein a sample in the set of trigger samples is one of: noise, a seed applied to a random number generator, and a data sample associated with at least one of the first second sets of real data samples and that has a given relationship to a classification boundary.
18 . The computer program product as described in claim 13 wherein the first and second permutations are packed with the first and second set of real data samples as a tile tensor.