Open-Set Radio Frequency Fingerprint Identification Method Based on Multi-Task Prototype Learning.

Area of Science:

• Cybersecurity
• Signal Processing
• Machine Learning

Background:

• Radio Frequency (RF) fingerprinting is crucial for Internet of Things (IoT) security.
• Existing methods often fail in real-world scenarios due to the 'closed-set' assumption, unable to detect novel devices.
• There is a need for robust open-set recognition methods for dynamic IoT environments.

Purpose of the Study:

• To propose an open-set radio frequency fingerprint identification (RFFI) method.
• To address the limitation of 'closed-set' assumptions in current RF fingerprinting techniques.
• To enhance the security and robustness of wireless authentication systems in the IoT.

Main Methods:

• Developed a Multi-Task Prototype Learning (MTPL) framework integrating an encoder, decoder, and classifier.
• Employed discriminative classification, generative reconstruction, and prototype clustering tasks within the deep network.
• Utilized Extreme Value Theory (EVT) to establish an adaptive open-set discrimination threshold based on prototype distances.

Main Results:

• The proposed MTPL method achieved a mean AUROC of 0.9918 on a real-world dataset with 16 Wi-Fi devices.
• Outperformed five mainstream open-set recognition methods, including SoftMax thresholding, OpenMax, and MLOSR.
• Demonstrated superior performance in identifying known and unknown devices, enhancing wireless authentication security.

Conclusions:

• The novel MTPL approach effectively handles open-set recognition challenges in RF fingerprinting.
• The integration of EVT provides a robust mechanism for adaptive thresholding.
• The method significantly improves the security and reliability of IoT wireless authentication systems.