Data-free Backdoor Attacks on Pre-trained Human Activity Sensing Models in IoT Systems Article

Zhao, T, Wang, X. (2026). Data-free Backdoor Attacks on Pre-trained Human Activity Sensing Models in IoT Systems . IEEE TRANSACTIONS ON MOBILE COMPUTING, 10.1109/TMC.2026.3745275

cited authors

  • Zhao, T; Wang, X

authors

abstract

  • Self-supervised learning (SSL) has emerged as a powerful deep learning paradigm for human activity recognition (HAR) systems. By leveraging large amounts of unlabeled data, SSL enables the development of pre-trained models (PTMs) that can be efficiently fine-tuned with limited labeled data for downstream HAR tasks, reducing labeling costs while improving generalization. Despite these advantages, the potential vulnerabilities of SSL-based PTMs in IoT sensing systems have not been sufficiently explored. This paper investigates data-free backdoor attacks on these PTMs, focusing on a practical scenario where attackers cannot access downstream task data. To realize these attacks, we design a set of triggers and predefined output representations (PORs). By mapping triggers to PORs through backdoor training, we can implant backdoor behaviors into the PTMs, thereby introducing vulnerabilities across different downstream sensing tasks without requiring prior knowledge. Extensive experiments across various sensing modalities, multiple datasets, and diverse PTM architectures demonstrate the broad effectiveness of the attack while preserving clean-task utility. Strict cross-dataset evaluations further show that the attack can transfer under domain shift and limited substitute-data availability. Furthermore, we evaluate several defense strategies, which prove ineffective in detecting or mitigating our attack. These findings highlight the need for robust security measures in IoT sensing PTMs.

publication date

  • January 1, 2026

published in

Digital Object Identifier (DOI)