A Robust NFT Assisted Knowledge Distillation Framework for Edge Computing

Nai Wang, Atul Sajjanhar, Yong Xiang, Longxiang Gao

Abstract: With the development and improvement in chip manufacturing and network communication, Internet of Things (IoT) have been addressing more and more popularity around these days. Due to the fact that the end devices in an IoT system can perform higher computational tasks, there are more and more IoT applications requiring on-device local training procedures. Hence, the concept of Knowledge Distillation is introduced to solve the on-device machine learning problem–each end device will receive a distilled light-weight student model from the comprehensive central teaching model. However, several security concerns need to be resolved before KD being put into industrial environments, including data integrity and robustness over external attacks. In this paper, we propose an NFT assisted KD framework, aiming at leveraging the blockchain features on data security to solve the intrinsic robustness defects in a naive KD architecture. Our major contributions can be concluded as following 1) the first NFT assisted KD framework (KD-NFT) which initializes the chance of NFT usages in scientific fields; 2) providing a two-dimension (vertical and horizontal) security over KD data vulnerability under attacks; and 3) a fail-over scheme when external poisoning happened, to recovering KD-NFT training process back to last-best status, by using NFT history full-traceable feature and providing automatic system robustness.