Toward Autonomous Physical Security Defenses Using Machine Learning

Abstract

The sheer increase in interconnected devices, reaching 50 B in 2025, makes it easier for adversaries to have direct access to the target system and perform physical attacks. This risk is exacerbated by the proliferation of Internet-of-Battlefield Things (IoBT) and increased reliance on the use of embedded devices in critical infrastructure and industrial control systems. Existing anti-tamper designs protect against limited forms of attacks and have deterministic tamper responses, which can undermine the availability of systems. Advancements in physical inspection techniques have enabled stealthier attacks. Therefore, there is a pressing need for more intelligent defenses that ensure a longer operational time while keeping up with the expected increase in the capabilities of adversaries. This study proposes to enhance existing physical protection methods by developing an intelligent anti-tamper using machine learning algorithms. It uses an analytic system capable of detecting and classifying multiple types of behaviors (e.g., normal operation conditions, known attack vectors, and anomalous behavior). The system also has a layered response mechanism and recovery scheme, which reduces false alarms and prolongs the operational time. An experimental platform was constructed and used for data collection and machine learning model training. This study also explored the impact of adversarial learning attacks on the proposed system and subsequently developed a countermeasure. The final prototype was capable of recognizing two types of normal operating conditions (sheltered and exposed environments) and four types of physical attacks. It also has adaptive response and recovery mechanisms.

Keywords

• Anti-tamper design
• internet of battlefield things (IoBT)
• machine learning algorithms
• adversarial machine learning
• physical attacks
• hardware security