The rise of Industry 4.0 and the transition toward Industry 5.0 have significantly increased the reliance on mission-critical smart systems, including smart manufacturing assets. However, this shift also introduces risks associated with adversarial manipulation and malicious exploitation of control channels, particularly in systems involving human, AI, and hybrid decision-making. As AI systems grow towards increasingly powerful ones, global concerns persist about the risks of their misuse or the dangers of granting them unrestrained autonomy. The DieHard framework offers a balanced solution by equipping mission-critical smart systems with limited, yet crucial, autonomy: the ability to detect misuse and ensure resilience against exploitation for malicious purposes. This approach enables AI systems to remain under human control while safeguarding their operations from manipulation or misuse, striking a vital balance between usability and security. This paper introduces a human-centric, responsible, and resilient autonomy framework that integrates anomaly detection, ethical constraints, and mission alignment. We propose a taxonomy of adversarial anomalies, outline the role of the Learning Entropy concept for anomaly detection across diverse environments, and provide simple proof-of-concept simulation. By maintaining the operator-in-the-loop and enforcing resilience under anomalous control, DieHard ensures robust, reliable, and secure operation of mission-critical smart systems, even in adversarial scenarios.
Terziyan et al. (Thu,) studied this question.