As deep learning models become more prevalent in\nsmart grid systems, ensuring their accuracy in tasks like identifying\nabnormal customer behavior is increasingly important. As its\nuse is increased in smart grids to detect energy theft, crafting adversarial\ndata by attackers to deceive the model to get the desired\noutput is also increased. Evasion attacks (EA) attempt to evade\ndetection by misclassifying input data during testing. The manipulation\nof data inputs is done so that it is not noticeable to humans\nbut can cause the machine learning (ML) model to produce incorrect\nresults. Electricity theft has become a major problem for utility\ncompanies that need to be dealt with effectively. Convolutional\nNeural Network (CNN) and AdaBoost hybrid model have been developed\nthat promise to detect electricity theft with high accuracy.\nHowever, this model is also vulnerable to evasion attacks that can\nrender it ineffective.<\/p>\n
In this paper, to make the detection system more robust, we\npresent a generative method to create evasion attacks against a\nhybrid model combining Convolutional Neural Network and\nAdaboost (CNN-Adaboost). Generated adversarial data from the\nproposed algorithm is crafted on the model to test its performance.\nOur proposed attack is validated with State Grid Corporation of\nChina (SGCC) dataset. We test the CNN-Adaboost energy theft\ndetection model and other models\u2019 performance under 5% and\n10% evasion attacks. Our findings reveal model performance degradation\nunder our proposed generative evasion attack ranging\nfrom 96.35% to 89.23%. With the defence mechanism, we successfully\nincreased adversarial accuracy by up to 97% and decreased\nthe attack success rate (ASR) by up to 3%. These adversaries are\nuseful for designing robust and secure machine learning models,\noffering an improved solution compared to previous work in this\narea. We tested the model with varying percentages of adversarial\ndata to analyze its behavior effectively. These adversaries are\nuseful for designing robust and secure ML models. The proposed\nattack and defence can be utilized to test energy theft detection\n(ETD) models in industrial and commercial settings.<\/p><\/h5>\n