BrInstFlip: A Universal Tool for Attacking DNN-Based Power Line Fault Detection Models
Yixin Jiang,Yanbo Xu,Zhihong Liang,Wenqian Xu,Tian Dong,Ximing Zhang,Yiwei Yang,Yan Meng,Haojin Zhu
DOI: https://doi.org/10.1109/iccc62479.2024.10681810
2024-01-01
Abstract:Deep learning neural network (DNN) models are currently experiencing significant success in domains like image classification. In the realm of power grids, there have been numerous attempts to leverage DNN models for fault diagnosis in power grid systems. However, they are also susceptible to various attacks, such as adversarial samples, model poisoning, backdoors, and other attack vectors. These existing attacks exhibit limitations to varying degrees, hampering their ability to execute generalized attacks across models. Designing a universal attack strategy for different DNN models poses challenges. In this paper, we present a novel attack strategy centered around runtime code fault injection, aiming to disrupt the inference capability of DNN architectural fault detection models in smart grid. Our approach meticulously identifies vulnerable bits within common code segments, such as underlying libraries, susceptible to functional bias through automated searches. Leveraging the DRAM Rowhammer Bug, we manipulate memory bits, inducing bias in the functionality of the affected library, consequently impacting all higher-level machine learning frameworks utilizing it. We conducted tests on the VGG-16, ResNet34, and ResNet50 models for the Vari-grip fault detection task using the InsPLAD dataset. Before the attack, the models achieved accuracies of 86.62%, 92.88%, and 97.12% on the test set, respectively. However, after the attack, their accuracies dropped to 0%, which can lead to severe consequences like failure of power grid repair. These results underscore the effectiveness and generalizability of our attack strategy.