University of Twente Student Theses

Login

Leveraging zero-knowledge succinct arguments of knowledge for efficient verification of outsourced training of artificial neural networks

Zande, M.J. van de (2019) Leveraging zero-knowledge succinct arguments of knowledge for efficient verification of outsourced training of artificial neural networks.

[img] PDF
402kB
Abstract:Incremental innovations have led to practical implementations of both Artificial Neural Networks and Succinct Non-interactive ARGuments, this thesis explores the practicality of using these SNARGs to verify outsourced training of an ANN. This training algorithm presents a particular case of Verifiable Computation as it relies on large quanti- ties of input data, floating point arithmetic and parallel computation. Decomposition of the core concepts ANNs and SNARGs to their elemen- tary building blocks allows the identification of imposed constraints mapped to the case of outsourced training and existing practical implementations of proof systems. The verdict separates the two com- putations and postulates how the Linear Probabilistically Checkable Proofs fit the inference computation and how the Interactive Oracle Proofs, specifically STARKs, fit the training computation. This work builds on the the works of Chabenne et al. and Ghodsi et al. that studied the verification of the inference computation and the work of Wu et al. that studies a similar training algorithm in the context of DIstributed Zero-Knowledge.
Item Type:Essay (Master)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Computer Science MSc (60300)
Link to this item:http://purl.utwente.nl/essays/79180
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page