Adversarial deep learning for energy management in buildings

Deep learning is a powerful means to classify and thus optimize Energy management in Buildings. Deep learning is effective especially when the training dataset has a reduced volume or when the test set changes at a higher frequency than the training set. Notwithstanding these favourable properties, the classification with deep learning could be distorted by an adversary who can be interested to alter the classification of the energy consumption. Several kinds of fraud could require this attack, as those aimed at energy theft. In this paper we will provide experimental implants where a dataset is tampered with in order to lead the classifier to acquire it as valid, while it contains samples attributable to energy thefts.