A Model-Independent Trojan Attack on Deep Learning-Based FDIA Detection in Smart Grid Protection Systems

Deep learning (DL) models have been proven to be highly effective in detecting false data injection attacks (FDIAs) in smart grids relying only on the measurements of the secured component. However, the vulnerability of these DL models to Trojan attacks poses significant security risks. This article introduces a model-independent Trojan attack (MITA) targeting DL models used for FDIA detection in critical smart grid components, such as transformer differential relays (TDRs). MITA injects a carefully crafted trigger signal into the TDR’s measurements, which evades detection by DL-based FDIA detection methods, causing FDIAs to be misclassified as faults. This leads to unnecessary tripping of the attacked TDR, potentially compromising grid reliability. Key features of MITA include model independence, enabling it to target diverse DL architectures without retraining, high success rates exceeding 99%, and stealthiness. The attack is validated against multiple DL models trained to detect FDIAs on TDRs using FDIA and fault scenarios generated in a realistic OPAL-RT environment. Our results show that MITA maintains the original model’s performance in the absence of the trigger while significantly misclassifying FDIAs as faults when the trigger is present. These results highlight the threat Trojan attacks can cause on industrial smart grid systems.