Industrial large models as carriers for malicious payloads: A fast and robust approach

Abstract

As artificial intelligence (AI) technologies are increasingly deployed in industrial domains, the use of pre-trained industrial large models (ILMs) has become widespread due to their cost-effectiveness and high performance across complex tasks. However, this growing reliance introduces new cybersecurity threats, particularly concerning the integrity of model parameters. Attackers are increasingly targeting AI models, exploiting vulnerabilities in the software supply chain, providing a covert means of executing novel cyberattacks, and posing significant security risks. Although antivirus software and intrusion detection systems are effective in protecting systems, the evolving nature of attack strategies, particularly the use of widely available AI models as carriers for malicious payloads, poses increasingly sophisticated security threats. Most existing embedding techniques struggle in ILM application scenarios, where payloads are inefficiently embedded and extracted, and are easily disrupted by fine-tuning. Meanwhile, the few robustness techniques always face limitations in areas such as low efficiency, model performance degradation, and the challenge of achieving a balance between efficiency and robustness. In this paper, we introduce FREEZER: Fast Redundant Exponent Embedding with Robustness (Robustness refers to the ability to preserve the embedded payload after full fine-tuning, while fast speed denotes its substantially faster payload embedding and extraction compared to prior approaches), a framework for significantly improving efficiency while maintaining robustness during the injection of malicious payloads into ILMs. FREEZER effectively addresses the challenge of extensive bit errors—defined as the bitwise discrepancy between the originally embedded malicious payload and the payload recovered by the prescribed extractor after full-parameter fine-tuning. Moreover, the infected models obtained using FREEZER exhibit no significant performance degradation. Experimental results show that FREEZER achieves a 20x faster injection speed and a 240x faster extraction speed compared to the current state-of-the-art (SOTA) method while maintaining high robustness. FREEZER raises awareness of this emerging threat and inspires the development of novel defenses against new forms of cyberattacks.