AI for Cyberattacks’ Detection in the Metaverse

Detecting malicious messages generated by cyberattacks is essential to provide security to Metaverse services. AI strategies based on autoencoders have been proven effective in revealing the presence of anomalous messages altered by malicious attacks. However, autoencoders’ algorithms are typically executed by HW accelerators, that are becoming increasingly prone to Soft Errors (SEs) occurring during their in-field operation. In this regard, we have recently shown that SEs affecting the Most Significant Bit of the weights of autoencoders for anomalous messages’ detection may significantly impact system’s availability. In this paper, we present a detailed analysis of the impact of SEs on the ability of autoencoders to detect anomalous messages generated by cyberattacks. We consider an autoencoder trained to discriminate credit card legal and illegal transactions. We analyze the diverse effects of SEs, as a function of the affected bits in the weights and layers of the autoencoder. This allows us to identify the bits in the weights and layers that, if affected by SEs, are more likely to affect the autoencoder ability to detect correctly the presence of anomalous messages generated by cyberattacks. Based on our analysis results, selective adoption of traditional fault tolerance techniques could then be employed for such identified autoencoder’s bits in the weights and layers only, thus enabling to protect autoencoders against SEs at low costs.