An Invisible Backdoor Attack on Cloud Computing in the Internet of Things

Cloud computing has significant advantages in IoT applications, such as high flexibility, scalability, and fast response. However, cloud computing relies on neural networks for processing and decision-making in big data scenarios, and there is a risk of implanted backdoors. A backdoor attack refers to an attack that leads to model misjudgment by implanting specific triggers into the original model during the training process of the neural network model. To study the security problem of cloud computing, this paper proposes an invisible backdoor attack scheme based on feature space similarity theory. The scheme first obtains the original triggers through a model inversion method and a set of random target samples. After that, the benign samples are segmented into feature regions by the attention mechanism, the triggers are added to the feature regions, and the generated poison samples are optimized to improve the stealthiness of the triggers and increase the poisoning efficiency. After expanding the poison dataset by image enhancement algorithm, the original model is retrained to generate the poison model. Ultimately, the experimental results show that the scheme achieves 97% attack success rate with 1% poison ratio in the automatic driving recognition scenario while ensuring the stealthiness of the trigger. Meanwhile, the poison model generated by this scheme can successfully escape detection by the defense algorithm, which improves the indistinguishability of the poison model.