Radio frequency fingerprint identification (RFFI) offers a promising physical-layer method to authenticate devices based on unique hardware impairments. However, existing RFFI systems use deep learning (DL) models that are resource- intensive. Training is particularly demanding, requiring repeated updates to a large number of parameters. In this paper, we introduce a quantum-assisted training (QAST) framework to address training inefficiencies in RFFI. QAST integrates a quantum neural network (QNN) with a mapping network to generate parameters for a classical DL model. This indirect training strategy substantially reduces the number of trainable parameters and the overall memory requirements compared to direct training of the DL model. We achieve this by introducing a multimodal mapping network that effectively learns the QNN output. This network generates multiple classical parameters from a shared quantum representation, thereby reducing qubit requirements and lowering the risk of barren-plateau-related trainability degradation. We also propose a new embedding method that reduces the size of the embedding matrix and yields a 15% to 30% reduction in training time. The tailored QAST framework trains the RFFI model while requiring only 10% of the original number of parameters while maintaining comparable accuracy, thereby substantially reducing memory and computational overhead and enabling efficient training or retraining in resource-limited environments.

Securing Internet of Things (IoT) devices presents increasing challenges due to their limited computational and energy resources. Radio Frequency Fingerprint Identification (RFFI) emerges as a promising authentication technique to identify wireless devices through hardware impairments. RFFI performance under low signal-to-noise ratio (SNR) scenarios is significantly degraded because the minute hardware features can be easily swamped in noise. In this paper, we leveraged the diffusion model to effectively restore the RFF under low SNR scenarios. Specifically, we trained a powerful noise predictor and tailored a noise removal algorithm to effectively reduce the noise level in the received signal and restore the device fingerprints. We used Wi-Fi as a case study and created a testbed involving 6 commercial off-the-shelf Wi-Fi dongles and a USRP N210 software-defined radio (SDR) platform. We conducted experimental evaluations on various SNR scenarios. The experimental results show that the proposed algorithm can improve the classification accuracy by up to 34.9%.

Deep learning-based Wi-Fi sensing has received massive interest thanks to the prevalence of Wi-Fi technology. While deep learning techniques provide promising results in Wi-Fi sensing, there are only very few studies on the vulnerabilities against Wi-Fi ensing. In this paper, we studied evasion attacks against deep learning-based Wi-Fi sensing and the countermeasure and conducted an extensive experimental evaluation using two publicly available datasets, namely SignFi and Widar. Accordingly, we proposed three white-box and two black-box attacks and revealed that even with an undetectable power change, evasion attacks can achieve a remarkable attack success rate (ASR) of 97.0% and 95.6% in white-box and black-box settings, respectively. These results highlight the urgent need for countermeasures against evasion attacks in Wi-Fi sensing systems. We introduced adversarial training and randomised smoothing, which notably improved the robustness of the Wi-Fi sensing model. The ASRs for white-box and black-box attacks were reduced to a minimum of around 6% and 2%, respectively. Moreover, randomised smoothing also introduced certifiable robustness, achieving 70.1% of samples certified for our model. The certification method provides an additional layer of reliability, ensuring that the model's performance remains consistent and predictable even under adversarial conditions.

Wi-Fi sensing can classify human activities because each activity causes unique changes to the channel state information (CSI). Existing WiFi sensing suffers from limited scalability as the system needs to be retrained whenever new classes are added, which causes overheads of data collection and retraining. Cross-domain sensing may fail because the mapping between activities and CSI variations is destroyed when a different environment or user (domain) is involved. This paper proposed a few-shot learning-based WiFi sensing system, named FewSense, which can recognise novel classes in unseen domains with only a few samples. Specifically, a feature extractor was pre-trained offline using the source domain data. When the system was applied in the target domain, a few samples were used to fine-tune the feature extractor for domain adaptation. Inference was made by computing the cosine similarity. FewSense can further boost the classification accuracy by collaboratively fusing inference from multiple receivers. We evaluated the performance of FewSense using three public datasets, i.e., SignFi, Widar, and Wiar. The results show that FewSense with five-shot learning recognised novel classes in unseen domains with an accuracy of 93.9%, 96.5%, and 82.7% on the SignFi, Widar, and Wiar datasets, respectively. Our collaborative sensing model improved system performance by an average of 29.2%.