SRAM and Generative Network-based Physical Fingerprinting for Trust Management in the Internet of Things

Recent advances in the Internet of Things (IoT), machine learning, and edge computing have led to the development of paradigms such as smart cities, smart grids, smart healthcare, and intelligent transportation systems as efficient and cost-effective solutions. Subsequently, there has also been an increase in the number of connected devices, ranging from high-power computers to low-power microcontrollers and sensors. The multi-layered and complex structure of the IoT creates a vast surface of vulnerabilities. Cyber threats such as proxy attacks are prevalent in ubiquitous resource-constrained IoT devices giving rise to the need for practical device fingerprinting algorithms. Existing works are based on network activity deep learning-based classification methods. However, attackers can mimic network activity, and classification models must be retrained on the data of new anomalies. This paper solves these issues by proposing a lightweight, and intelligent physical fingerprinting algorithm and the corresponding mutual authentication protocol using initial power-up Static Random Access Memory (SRAM) states and generative networks. A generative network is trained to reconstruct SRAM fingerprints of an authorized device during the registration phase and an anomaly threshold is selected. The proposed technique reliably fingerprints registered devices, and can detect proxy devices of identical architectures from the same and different manufacturers with high accuracy. Since the algorithm does not require attacker data during the training phase, it is a self-sufficient and low-cost solution. The method has a latency of 2.114 seconds per device and security analysis of the proposed protocol proves its security against proxy and de-synchronization attacks.