Impact of Soft Errors on High Performance Autoencoders for Cyberattack Detection

Detecting anomalous messages generated by cyberattacks is essential for IoT based critical applications (e.g., finance, healthcare, manufacturing, etc.), in order to guarantee high levels of security and availability. Autoencoder-based anomaly detection has been proven effective in detecting the presence of messages altered by malicious attacks. To guarantee low detection latency, autoencoders’ algorithms are typically executed by HW accelerators implemented by nanotechnology. However, such HW accelerators are susceptible to soft errors (SEs), that may occur during their in-field operation. In this paper, we analyze the impact of SEs on the effectiveness of autoencoders in detecting anomalous messages generated by cyberattacks. As an example, we consider autoencoders trained to discriminate credit card legal and illegal transactions. We show that SEs do not decrease system’s security, but may reduce system’s availability of the 70%, so that proper solutions to avoid the detrimental impact of SEs on system’s availability should be devised for autoencoders implemented by high performance HW accelerators.