Tweak: Towards Portable Deep Learning Models for Domain-Agnostic LoRa Device Authentication

TL;DR

This paper introduces Tweak, a lightweight domain adaptation technique for deep learning-based device fingerprinting that enables models trained in one domain to perform well in another with minimal additional data, enhancing portability and efficiency.

Contribution

Tweak leverages metric learning and calibration to adapt models across domains without retraining, suitable for resource-constrained IoT environments.

Findings

Tweak improves device identification accuracy across different domains.

The method requires only a small amount of target domain data.

Tweak is computationally lightweight and suitable for IoT networks.

Abstract

Deep learning based device fingerprinting has emerged as a key method of identifying and authenticating devices solely via their captured RF transmissions. Conventional approaches are not portable to different domains in that if a model is trained on data from one domain, it will not perform well on data from a different but related domain. Examples of such domains include the receiver hardware used for collecting the data, the day/time on which data was captured, and the protocol configuration of devices. This work proposes Tweak, a technique that, using metric learning and a calibration process, enables a model trained with data from one domain to perform well on data from another domain. This process is accomplished with only a small amount of training data from the target domain and without changing the weights of the model, which makes the technique computationally lightweight and thus suitable for resource-limited IoT networks. This work evaluates the effectiveness of Tweak vis-a-vis its ability to identify IoT devices using a testbed of real LoRa-enabled devices under various scenarios. The results of this evaluation show that Tweak is viable and especially useful for networks with limited computational resources and applications with time-sensitive missions.