부산대학교 도서관

Data-driven electricity theft detectors use consumer reported power consumption measurement to detect energy theft. Most of the detection scheme incorporate machine learning algorithms to classify unknown theft. The correct labelling of the training data is a common implicit assumption in such detectors. Unfortunately, these detectors are vulnerable against data poisoning attacks that assume false labels during training. However, the difficulty of specific and accurate features selection arise due to lack of training data set, existing schemes are unable to detect unknown attacks effectively. To enhance the detectors' robustness against unknown data attacks, we propose a theft detection scheme based on label propagation based semisupervised learning algorithm by using small amount of labeled dataset. The scheme is achieve high accuracy than other existing algorithms. The proposed scheme than compared with the other supervised schemes.