A Data-free Backdoor Injection Approach in Neural Networks

Recently, the backdoor attack on deep neural networks (DNNs) has been extensively studied, which causes the back-doored models to behave well on benign samples, whereas performing maliciously on controlled samples (with triggers attached). Almost all existing backdoor attacks require access to the original training/testing dataset or data relevant to the main task to inject backdoors into the target models, which is unrealistic in many scenarios, e.g., private training data. In this paper, we propose a novel backdoor injection approach in a "data-free" manner1. We collect substitute data irrelevant to the main task and reduce its volume by filtering out redundant samples to improve the efficiency of backdoor injection. We design a novel loss function for fine-tuning the original model into the backdoored one using the substitute data, and optimize the fine-tuning to balance the backdoor injection and the performance on the main task. We conduct extensive experiments on various deep learning scenarios, e.g., image classification, text classification, tabular classification, image generation, and multimodal, using different models, e.g., Convolutional Neural Networks (CNNs), Autoencoders, Transformer models, Tabular models, as well as Multimodal DNNs. The evaluation results demonstrate that our data-free backdoor injection approach can efficiently embed backdoors with a nearly 100% attack success rate, incurring an acceptable performance downgrade on the main task.