Exploring Adversarial Attacks on the MaSTer Truncation Protocol

At CANS 2024, Zbudila et al. presented MaSTer, a maliciously secure multi-party computation protocol for truncation.It allows adversaries to manipulate outputs with a bounded additive error while avoiding detection with a certain probability.In this work, we analyse the broader implications of adversarial exploitation in probabilistic truncation protocols, specifically in relation to MaSTer.We propose three attack strategies aimed at inducing misclassification in deep neural network (DNN) inference.Our empirical evaluation across multiple datasets demonstrates that while adversarial influence remains negligible under realistic constraints, certain configurations and network architectures exhibit increased vulnerability.By improving the understanding of the risks associated with probabilistic truncation protocols in privacy-preserving machine learning, our work demonstrates that the MaSTer protocol is robust in realistic settings.