Detecting Stuxnet‐like data integrity attacks
Tuuli Siiskonen, Mika Rantonen · Security and Privacy · 2020
Abstract This work suggests and tests a solution for detecting loss of integrity with means of prediction in distributed, low‐power computing environment or when collecting relational data disconnected from original data source to data ecosystems. With graph convolution network and variational auto‐encoder, the proposed solution is able to detect data integrity attacks based on link prediction for datasets from various data sources without common data model. The solution has in‐built mechanism to prevent poisoning the detection capability, and it is resource efficient since it does not require extensive storage capability. Due to its programmatic nature and high ability to integrate, unsupervised learning‐based solution fits well in embedded system ecosystems' control functions where automated decisions are taken.