Detecting and defending against adversarial attacks in decentralized machine learning systems
View Patent ↗A system and a method for detecting and defending against adversarial attacks in decentralized learning models are described. The method comprises obtaining a learning parameter for determining a reference cryptographic hash value and a similarity between the data processing nodes ( 102 ). A cryptographic hash value is determined for each data processing node ( 102 ) based on the learning parameter. The trust score of each data processing node ( 102 ) is updated based on matching of the cryptographic hash value with the reference cryptographic hash value. The learning parameter of each data processing node ( 102 ) is merged to obtain a merged learning parameter based on the trust score. The merged learning parameter is provided to the data processing nodes ( 102 ) to be used for training the machine learning models.
1. A method comprising:
obtaining, by each of a plurality of data processing nodes ( 102 ), a learning parameter associated with training of a plurality of machine learning models, wherein the plurality of data processing nodes ( 102 ) are enrolled with a decentralized blockchain network ( 100 );
determining, by each of the plurality of data processing nodes ( 102 ), a reference cryptographic hash value and a similarity between learnings of the plurality of machine learning models based on the learning parameter, wherein the reference cryptographic hash value refers to an identity of the corresponding data processing node;
providing, by each of the plurality of data processing nodes ( 102 ), the reference cryptographic hash value, the similarity, and the learning parameter to a leader node ( 102 - 1 ), wherein the leader node ( 102 - 1 ) is one of the plurality of data processing nodes ( 102 );
determining, by the leader node ( 102 - 1 ), a trust score of each of the plurality of data processing nodes ( 102 ) based on the similarity, wherein the trust score of a data processing node indicates genuinity of the data processing node;
determining, by the leader node ( 102 - 1 ), a cryptographic hash value for each of the plurality of data processing nodes ( 102 ) based on the learning parameter of each of the plurality of data processing nodes ( 102 );
updating, by the leader node ( 102 - 1 ), the trust score of each of the plurality of data processing nodes ( 102 ) based on matching of the cryptographic hash value with the reference cryptographic hash value;
merging, by the leader node ( 102 - 1 ), the learning parameter of each of the plurality of data processing nodes ( 102 ) to obtain a merged learning parameter, wherein the learning parameter of each of the plurality of data processing nodes ( 102 ) are merged based on the trust score of each of the plurality of data processing nodes ( 102 ); and
providing, by the leader node ( 102 - 1 ), the merged learning parameter to the plurality of data processing nodes ( 102 ) to be used for training the plurality of machine learning models.
2. The method as claimed in claim 1 , further comprising:
determining, by the leader node ( 102 - 1 ), a weightage of each of the plurality of data processing nodes ( 102 ) based on the trust score, wherein the learning parameter of each of the plurality of data processing nodes ( 102 ) are merged based on the weightage of each of the plurality of data processing nodes ( 102 ).
3. The method as claimed in claim 1 , wherein the similarity is determined by:
determining, by each of the plurality of data processing nodes ( 102 ), a cosine similarity value between gradients of top layers of the reference cryptographic hash value;
performing, by each of the plurality of data processing nodes ( 102 ), dimensionality reduction on the cosine similarity values to obtain a compressed cosine similarity; and
determining, by each of the plurality of data processing nodes ( 102 ), centroid of each compressed similarity, thereby determining the similarity.
4. The method as claimed in claim 3 , wherein a value of the similarity close to 1 indicates a genuine data processing node and a value of the similarity close to −1 indicates a malicious data processing node.
5. The method as claimed in claim 1 , further comprising:
publishing, by the leader node ( 102 - 1 ), the merged learning parameter to a distributed ledger ( 208 ); and
training, by each of the plurality of data processing nodes ( 102 ), the plurality of machine learning models based on the merged learning parameter.
6. The method as claimed in claim 1 , wherein the merged learning parameter is obtained by:
collecting, by the leader node ( 102 - 1 ), the similarity between the plurality of data processing nodes ( 102 );
clustering, by the leader node ( 102 - 1 ), the plurality of data processing nodes ( 102 ) to form a node cluster based on the similarity;
identifying, by the leader node ( 102 - 1 ), one or more malicious data processing nodes from the plurality of data processing nodes ( 102 ), wherein the one or more malicious data processing nodes are associated with a value of similarity lesser than a pre-defined baseline cryptographic hash value; and
merging, by the leader node ( 102 - 1 ), the learning parameter obtained from each of the plurality of data processing nodes ( 102 ) excluding the one or more malicious data processing nodes.
7. The method as claimed in claim 6 , wherein the pre-defined baseline cryptographic hash value is obtained from a cryptographic hash value of a merged learning parameter obtained from previous learning of the plurality of machine learning models.
8. The method as claimed in claim 6 , wherein details of each of the one or more malicious data processing nodes are published on a distributed ledger ( 208 ).
9. The method as claimed in claim 1 , wherein a data processing node having the trust score below a predefined threshold value is excluded for election as the leader node.
10. A system comprising:
a plurality of data processing nodes ( 102 ) enrolled in a decentralized blockchain network ( 100 ), wherein each of the plurality of data processing nodes ( 102 ) comprises a corresponding machine learning model of a plurality of machine learning models and is programmed to:
obtain a learning parameter associated with training of the corresponding machine learning model, wherein the plurality of data processing nodes ( 102 ) are enrolled with a decentralized blockchain network ( 100 );
determine a reference cryptographic hash value and a similarity between learnings of the plurality of machine learning models based on the learning parameter, wherein the reference cryptographic hash value refers to an identity of the corresponding data processing node; and
provide the reference cryptographic hash value, the similarity, and the learning parameter to a leader node ( 102 - 1 ) elected from the plurality of data processing nodes ( 102 ), wherein the leader node is configured to:
determine a trust score of each of the plurality of data processing nodes ( 102 ) based on the similarity, wherein the trust score of a data processing node indicates genuinity of the data processing node;
determine a cryptographic hash value for each of the plurality of processing nodes ( 102 ) based on the learning parameter of each of the plurality of data processing nodes ( 102 );
update the trust score of each of the plurality of data processing nodes ( 102 ) based on matching of the cryptographic hash value with the reference cryptographic hash value;
merge the learning parameter of each of the plurality of data processing nodes ( 102 ) to obtain a merged learning parameter, wherein the learning parameter of each of the plurality of data processing nodes ( 102 ) is merged based on the trust score of each of the plurality of data processing nodes ( 102 ); and
provide the merged learning parameter to the plurality of data processing nodes ( 102 ) to be used for training the corresponding machine learning models.
11. The system as claimed in claim 10 , wherein the similarity is determined by:
determining, by each of the plurality of data processing nodes ( 102 ), a cosine similarity value between gradients of top layers of the reference cryptographic hash value;
performing, by each of the plurality of data processing nodes ( 102 ), dimensionality reduction on the cosine similarity values to obtain a compressed cosine similarity; and
determining, by each of the plurality of data processing nodes ( 102 ), centroid of each compressed similarity, thereby determining the similarity.
12. The system as claimed in claim 11 , wherein a value of the similarity close to 1 indicates a genuine data processing node and a value of the similarity close to −1 indicates a malicious data processing node.
13. The system as claimed in claim 10 , wherein the leader node ( 102 - 1 ) is configured to publish the merged learning parameter to a distributed ledger ( 208 ), and each of the plurality of data processing nodes ( 102 ) is configured to train the plurality of machine learning models based on the merged learning parameter.
14. The system as claimed in claim 10 , wherein the merged learning parameter is obtained by:
collecting, by the leader node ( 102 - 1 ), the similarity between the plurality of data processing nodes ( 102 );
clustering, by the leader node ( 102 - 1 ), the plurality of data processing nodes ( 102 ) to form a node cluster based on the similarity;
identifying, by the leader node ( 102 - 1 ), one or more malicious data processing nodes from the plurality of data processing nodes ( 102 ), wherein the one or more malicious data processing nodes are associated with a value of similarity lesser than a pre-defined baseline cryptographic hash value; and
merging, by the leader node ( 102 - 1 ), the learning parameter obtained from each of the plurality of data processing nodes ( 102 ) excluding the one or more malicious data processing nodes.
15. The system as claimed in claim 10 , wherein the pre-defined baseline cryptographic hash value is obtained from a cryptographic hash value of a merged learning parameter obtained from previous learning of the plurality of machine learning models.
16. The system as claimed in claim 10 , wherein details of each of the one or more malicious data processing nodes are published on a distributed ledger ( 208 ).
17. The system as claimed in claim 10 , wherein the learning parameter is obtained by a secured container of each of the plurality of data processing nodes ( 102 ) from a corresponding software container configured for training of the machine learning model.
18. The system as claimed in claim 17 , wherein the secured container provides the merged learning parameter to the corresponding software container for training of the machine learning model locally.
19. The system as claimed in claim 10 , wherein a data processing node having the trust score below a predefined threshold value is excluded for election as the leader node.
20. A non-transitory computer-readable medium storing a program including instructions that, when executed by a processor, causes a system to execute the method of claim 1 .