A safe and practical distributed model collaborative training method and system

Abstract

The application discloses a kind of safe and practical distributed model collaborative training methods.Method includes: through server and management component cooperation, for each participant configuration hardware isolation environment, and by server to each data holder, i.e.to each participant distributes identity mark and root key, to establish the key hierarchy system across entity;Utilize management component in multiple mutually distrustful participants between dynamic switching GPU connection, each participant is based on the key hierarchy system independently derived session key, and private data is locally encrypted and written into GPU secure memory, while synchronizing GPU driver context between participants;GPU migrates back to server after completing the access of all participants, and the session key of each participant is reproduced by the server, the data is decrypted in the trusted execution environment TEE and the model joint training is performed.The application realizes the safe collaborative training of multi-party data under the premise of realizing the continuous access of GPU to data holder and avoiding the real-time leading key update of server, with the advantages of simple process, efficient key management, strong security and the like.

Images