Hijacking Attacks against Neural Networks by Analyzing Training Data

Backdoors and adversarial examples are the two primary threats currently faced by deep neural networks (DNNs). Both attacks attempt to hijack the model behaviors with unintended outputs by introducing (small) perturbations to the inputs. Backdoor attacks, despite the high success rates, often require a strong assumption, which is not always easy to achieve in reality. Adversarial example attacks, which put relatively weaker assumptions on attackers, often demand high computational resources, yet do not always yield satisfactory success rates when attacking mainstream black-box models in the real world. These limitations motivate the following research question: can model hijacking be achieved more simply, with a higher attack success rate and more reasonable assumptions? In this paper, we propose CleanSheet, a new model hijacking attack that obtains the high performance of backdoor attacks without requiring the adversary to tamper with the model training process. CleanSheet exploits vulnerabilities in DNNs stemming from the training data. Specifically, our key idea is to treat part of the clean training data of the target model as "poisoned data," and capture the characteristics of these data that are more sensitive to the model (typically called robust features) to construct "triggers." These triggers can be added to any input example to mislead the target model, similar to backdoor attacks. We validate the effectiveness of CleanSheet through extensive experiments on 5 datasets, 79 normally trained models, 68 pruned models, and 39 defensive models. Results show that CleanSheet exhibits performance comparable to state-of-the-art backdoor attacks, achieving an average attack success rate (ASR) of 97.5 respectively. Furthermore, CleanSheet consistently maintains a high ASR, when confronted with various mainstream backdoor defenses.