The multi-level threats attribution in conventional Internet of Things (IoT) environments is mainly achieved through the classification of network data relationships, which overlooks the similarity between conventional data and threat data, leading to a large number of false positives in the attribution results. In response to this, a multi-level threats attribution method for IoT environments based on deep learning and genetic algorithms is proposed. A deep learning neural network is established to identify threat data in the IoT environment, and batch normalization operations are added to separate conventional data from threat data, extracting features of multi-level threats data. Genetic algorithms are applied to obtain the optimal individual, achieving the initial node attribution and positioning of threat data. Experimental results show that the attribution results obtained using the proposed method have fewer false positives and are more accurate, meeting the practical needs for the security maintenance of IoT environments.