Abstract: When providing a service that utilizes a machine learning model, the countermeasures against cyber-attacks are required. The model extraction attack is one of the attacks, in which an attacker attempts to replicate the model by obtaining a large number of input-output pairs. While a defense using Proof of Work has already been proposed, an attacker can still conduct model extraction attacks by increasing their computational power. Moreover, this approach leads to unnecessary energy consumption and might not be environmentally friendly. In this paper, the defense method using Proof of Spacetime instead of Proof of Work is proposed to reduce the energy consumption. The Proof of Spacetime is a method to impose spatial and temporal costs on the users of the service. While the Proof of Work makes a user to calculate until permission is granted, the Proof of Spacetime makes a user to keep a result of calculation, so the energy consumption is reduced. Through computer simulations, it was found that systems with Proof of Spacetime, compared to those with Proof of Work, impose 0.79 times the power consumption and 1.07 times the temporal cost on the attackers, while 0.73 times and 0.64 times on the non-attackers. Therefore, the system with Proof of Spacetime can prevent model extraction attacks with lower energy consumption.

Tatsuki Fukuda, “Proof of Spacetime as a Defensive Technique Against Model Extraction Attacks” International Journal of Advanced Computer Science and Applications(IJACSA), 14(6), 2023. http://dx.doi.org/10.14569/IJACSA.2023.0140611

@article{Fukuda2023,
title = {Proof of Spacetime as a Defensive Technique Against Model Extraction Attacks},
journal = {International Journal of Advanced Computer Science and Applications},
doi = {10.14569/IJACSA.2023.0140611},
url = {http://dx.doi.org/10.14569/IJACSA.2023.0140611},
year = {2023},
publisher = {The Science and Information Organization},
volume = {14},
number = {6},
author = {Tatsuki Fukuda}
}