Data reproducibility using blockchains
Data verification in federate learning is faster and simpler. As artificial intelligence grows in usage, data verification is needed to prove custody and/or control. Electronic data representing an original version of training data may be hashed to generate one or more digital signatures. The digital signatures may then be incorporated into one or more blockchains for historical documentation. Any auditor may then quickly verify and/or reproduce the training data using the digital signatures. For example, a current version of the training data may be hashed and compared to the digital signatures generated from the current version of the training data. If the digital signatures match, then the training data has not changed since its creation. However, if the digital signatures do not match, then the training data has changed since its creation. The auditor may thus flag the training data for additional investigation and scrutiny.
1. A memory device storing instructions that when executed cause a hardware processor to perform operations, the operations comprising:
receiving local updates sent from a mobile device, the local updates generated by a federated learning model executed by the mobile device;
generating a learning modification to the federated learning model based on the local updates sent from the mobile device;
splitting the local updates into multiple shares via a secret sharing algorithm;
generating cryptographic hash values by hashing the multiple shares using a cryptographic hashing algorithm;
determining a number N B of different blockchains for a distribution of the cryptographic hash values, the number N B of the different blockchains based on a total number N S of the multiple shares according to a ratio of N S /N B having a maximum value;
distributing the cryptographic hash values via the different blockchains, wherein the different blockchains comprise a master blockchain dedicated to the mobile device and a sub-blockchain dedicated to the federated learning model;
retrieving verification hash values generated by the hashing of current versions of the local updates associated with the federated learning model using the cryptographic hashing algorithm;
comparing the cryptographic hash values distributed via the different blockchains to the verification hash values generated by the hashing of the current versions of the local updates; and
verifying that the current versions of the local updates are authentic in response to the verification hash values satisfying the cryptographic hash values distributed via the different blockchains.
2. The memory device of claim 1 , wherein the operations further comprise determining a ratio N S /N B based on the number N B of the different blockchains and a total number N S of the multiple shares split via the secret sharing algorithm.
3. The memory device of claim 1 , wherein the operations further comprise determining a number N R of recipients of the different blockchains.
4. The memory device of claim 3 , wherein the operations further comprise determining the number N B of the different blockchains based on the number N R of recipients of the different blockchains according to a ratio of N R /N B having a maximum value.
5. A system, comprising:
a hardware processor; and
a memory device, the memory device storing instructions, the instructions when executed causing the hardware processor to perform operations, the operations comprising:
receiving local updates generated by a federated learning model executed by a mobile device;
generating a learning modification to the federated learning model based on the local updates;
splitting the local updates into multiple shares via a secret sharing algorithm;
generating cryptographic hash values by hashing the multiple shares using a cryptographic hashing algorithm;
determining a number N B of different blockchains for a distribution of the cryptographic hash values, the number N B of the different blockchains based on a total number N S of the multiple shares according to a ratio of N S /N B having a maximum value;
distributing the cryptographic hash values via a master blockchain of the different blockchains dedicated to the mobile device and via a sub-blockchain of the different blockchains dedicated to the federated learning model;
retrieving verification hash values generated by the hashing of current versions of the local updates associated with the federated learning model using the cryptographic hashing algorithm;
comparing the cryptographic hash values distributed via the different blockchains to the verification hash values generated by the hashing of the current versions of the local updates; and
verifying that the current versions of the local updates are authentic in response to the verification hash values satisfying the cryptographic hash values distributed via the different blockchains.
6. The system of claim 5 , wherein the operations further comprise determining a number N R of recipients of the different blockchains.
7. The system of claim 5 , wherein the operations further comprise determining the number N B of the different blockchains based on the number N R of recipients of the different blockchains according to a ratio of N R /N B having a maximum value.
8. The system of claim 5 , wherein the operations further comprise determining that the current versions of the local updates are inauthentic in response to the verification hash values failing to satisfy the cryptographic hash values.
9. The system of claim 5 , wherein the operations further comprise receiving a noun identifier that identifies the federated learning model.
10. The system of claim 5 , wherein the operations further comprise receiving a noun identifier that identifies the mobile device.
11. A method, comprising:
receiving, by a server, local updates sent from a mobile device, the local updates generated by a federated learning model executed by the mobile device;
generating, by the server, a learning modification to the federated learning model based on the local updates sent from the mobile device;
splitting, by the server, the local updates into multiple shares via a secret sharing algorithm;
generating, by the server, cryptographic hash values by hashing the multiple shares using a cryptographic hashing algorithm;
determining, by the server, a number N B of different blockchains for a distribution of the cryptographic hash values, the number N B of the different blockchains based on a total number N S of the multiple shares according to a ratio of N S /N B having a maximum value;
distributing, by the server, the cryptographic hash values via a master blockchain of the different blockchains dedicated to the mobile device and via a sub-blockchain of the different blockchains dedicated to the federated learning model;
retrieving, by the server, verification hash values generated by the hashing of current versions of the local updates associated with the federated learning model using the cryptographic hashing algorithm;
comparing, by the server, the cryptographic hash values to the verification hash values generated by the hashing of the current versions of the local updates; and
verifying, by the server, that the current versions of the local updates are authentic in response to the verification hash values satisfying the cryptographic hash values.
12. The method of claim 11 , further comprising determining a number N R of recipients of the different blockchains.
13. The method of claim 12 , further comprising determining the number N B of the different blockchains based on the number N R of recipients of the different blockchains according to a ratio of N R /N B .
14. The method of claim 11 , further comprising determining that the current versions of the local updates are inauthentic in response to the verification hash values failing to satisfy the cryptographic hash values.
15. The method of claim 11 , further comprising receiving a noun identifier that identifies the federated learning model.
16. The method of claim 11 , further comprising receiving a noun identifier that identifies the mobile device.