TriggerCraft: A Framework for Enabling Scalable Physical Backdoor Dataset Generation with Generative Models

Sze Jue Yang, Chinh Duc La, Quang H Nguyen, Eugene Bagdasarian, Kok-Seng Wong, Chee Seng Chan, Khoa D Doan

Keywords: Backdoor attacks, Physical backdoor attacks, Datasets

TL;DR: TriggerCraft is a generative framework that automates physical backdoor dataset creation, enabling realistic attack simulation and research without real-world setup.

Abstract: Backdoor attacks, representing an emerging threat to the integrity of deep neural networks have received significant attention due to their ability to compromise deep learning systems covertly. While numerous backdoor attacks occur within the digital realm, their practical implementation in real-world prediction systems remains limited and vulnerable to disturbances in the physical world. Consequently, this limitation has led to the development of physical backdoors, where trigger objects manifest as physical entities within the real world. However, creating a requisite dataset to study physical backdoors is a daunting task. This hinders backdoor researchers and practitioners from studying such backdoors, leading to stagnant research progresses. This paper presents a framework namely as TriggerCraft that empowers researchers to effortlessly create a massive physical backdoor dataset with generative modeling. Particularly, TriggerCraft involves three automatic modules: suggesting the suitable physical triggers, generating the poisoned candidate samples (either by synthesizing new samples or editing existing clean samples), and finally selecting only the most plausible ones. As such, it effectively mitigates the perceived complexity associated with creating a physical backdoor dataset, converting it from a daunting task into an attainable objective. Extensive experiment results show that datasets created by TriggerCraft achieve similar observations with the real physical world counterparts in terms of both attacks and defenses, exhibiting similar properties compared to previous physical backdoor studies. This paper offers researchers a valuable toolkit for advancing the frontier of physical backdoors, all within the confines of their laboratories.

Primary Area: Social and economic aspects of machine learning (e.g., fairness, interpretability, human-AI interaction, privacy, safety, strategic behavior)

Submission Number: 21812