Deep Learning-Enabled Zero-Touch Device Identification: Mitigating the Impact of Channel Variability Through MIMO Diversity

Abstract: Deep learning-enabled device fingerprinting has proven efficient in enabling automated identification and authentication of transmitting devices. It does so by leveraging the transmitters' unique features that are inherent to hardware impairments caused during manufacturing to extract device-specific signatures that can be exploited to uniquely distinguish and separate between (identical) devices. Though shown to achieve promising performances, hardware fingerprinting approaches are known to suffer greatly when the training data and the testing data are generated under different channels conditions that often change when time and/or location changes. To the best of our knowledge, this work is the first to use MIMO diversity to mitigate the impact of channel variability and provide a channel-resilient device identification over flat fading channels. Specifically, we show that MIMO can increase the device classification accuracy by up to about $50\%$ when model training and testing are done over the same channel and by up to about $70\%$ when training and testing are done over different fading channels.

• The paper demonstrates that integrating MIMO diversity with deep learning significantly enhances RF device identification, with up to 70% improvement under mismatched channel gains.
• It employs full and partial blind channel estimation to recover device-specific signal artifacts, effectively mitigating channel fading and hardware impairments.
• The study highlights that while MIMO approaches excel in static or low-mobility conditions, high Doppler shifts present challenges for further robustness.

Deep Learning-Enabled Zero-Touch Device Identification: Mitigating Channel Variability with MIMO Diversity

Introduction

Deep learning-based RF device fingerprinting employs innate, hardware-induced signal artifacts to provide zero-touch identification and authentication for wireless devices. However, the utility of such approaches is undermined by their pronounced sensitivity to channel variability—classification performance degrades significantly when models are exposed to previously unseen propagation effects, a typical occurrence in dynamic wireless environments. The examined paper addresses this challenge by systematically studying the application of MIMO diversity for channel-resilient device fingerprinting under Rayleigh flat fading and rigorously quantifying the achievable gains over conventional SISO models (2306.07878).

Figure 1: Overview of the proposed framework. Dashed arrows show the varying channel pipeline. Red arrows show the traditional SISO approach pipeline.

MIMO Diversity for Fingerprinting Robustness

The framework leverages MIMO’s spatial diversity, particularly through the use of orthogonal space-time block codes (OSTBC), to enhance SNR and exploit both transmit and receive diversity. After MIMO transmission, blind channel estimation techniques—both full and partial—are employed to recover the transmitted signals from the composite received stream. The key element is that reconstructed signals correspond more directly to device-induced distortions, being less affected by random channel fading. This preserves the discriminative features exploited by downstream CNN classifiers.

The signal acquisition pipeline comprises MATLAB-based simulation of IEEE 802.11ac waveforms from 20 devices, parameterized with varied degrees of hardware impairments (IQ imbalance, phase noise, frequency offset, DC offset). The learning model utilizes a two-layer CNN, ingesting split $2 \times 160$ I/Q vectors from the channel-equalized signals.

Evaluation: Impact of Channel and Device Diversity

Path Gain Variability Experiments

Extensive experiments analyze model accuracy and resiliency under mismatched train/test channel conditions, varying average path gain (APG) over a broad dB range. Standard SISO-CNN pipelines demonstrate severe accuracy collapse (from 56% to 22%) when faced with substantial path gain mismatch, highlighting profound channel dependency.

By contrast, MIMO-enabled pipelines with full blind channel estimation achieve up to 70% absolute improvement in classification accuracy under mismatched conditions, sustaining above 90% accuracy across a wide range of test APG values. Partial blind estimation methods also provide significant gains, yet lag behind the full approach due to unavoidable residual ambiguities.

Figure 2: The accuracy and robustness of MIMO-enabled models compared to SISO under varying APG (Average Path Gain) settings and the resultant accuracy gap.

Doppler Shift and Device Impairment Diversity

Variable maximum Doppler shift (MDS) is introduced to model receiver mobility. While MIMO-enabled methods substantially outperform SISO under static or low-mobility conditions, all approaches experience pronounced degradation for high MDS scenarios; this is especially evident as device cardinality increases, suggesting that Doppler-induced channel evolution is a limiting factor for all examined methods.

Figure 3: Testing accuracy under various Doppler shifts, illustrating the limitations of MIMO diversity in high-mobility environments.

The intensity of device hardware impairment is also evaluated. As the impairment variance increases (making device fingerprints more separable), the accuracy of MIMO-enabled models improves markedly (up to 90% for 20 devices under high impairment), whereas SISO models remain near-random. This underlines the effectiveness of spatial-processing-based mitigation for tackling signal similarity due to homogeneous hardware.

Figure 4: The effect of IQ imbalance intensity on classification accuracy for SISO and MIMO approaches.

Strong Numerical Results and Claims

• MIMO diversity increases classification accuracy by up to 50% in matched channel conditions and up to 70% in mismatched conditions.
• Full blind channel estimation consistently outperforms partial estimation, especially as channel conditions diverge between training and testing.
• For low impairment and large device sets (20 devices), only MIMO-enabled methods provide meaningful classification, while SISO approaches degrade to near-chance accuracy.

Theoretical and Practical Implications

The use of MIMO spatial diversity in device fingerprinting demonstrates that classical channel-induced artefacts can be effectively separated from device-specific signatures via advanced channel estimation. This supports the design of real-world, scalable RF authentication systems robust to environmental variation—critical for large-scale IoT deployments and in-the-wild device tracking scenarios.

However, the inability of these methods to compensate for rapid temporal channel variation (Doppler) identifies a fundamental open challenge. Adaptations, such as explicit channel tracking or fusion with temporal models, may be necessary to ensure robust operation in mobile or crowded environments.

The evaluation clearly demonstrates that MIMO strategies and improved channel compensation—not merely neural network complexity or dataset augmentation—are pivotal for closing generalization gaps in RF fingerprinting systems.

Future Research Directions

• Doppler Resilience: Investigating cross-domain adaptation techniques to mitigate device misclassification under high-mobility.
• Scalability: Exploring advanced network architectures and leveraging further spatial/temporal diversity to maintain accuracy as the number of devices increases.
• Aging and Environmental Drift: Modeling and compensating for long-term changes in device fingerprints due to aging and environmental stressors.
• Adversarial Robustness: Assessing attack surfaces relating to neural-based fingerprinting, including effective defenses against crafted interference or adversarial examples.

Conclusion

This work rigorously quantifies the benefits of integrating MIMO diversity with deep learning-based device identification, establishing that channel-resilient RF fingerprinting requires spatial diversity and explicit channel compensation. While MIMO-enabled models achieve robust performance across a range of impairment and channel conditions, Doppler effects and device scalability remain critical challenges for subsequent research. These insights underscore a shift toward holistic physical-layer solutions underpinning next-generation wireless security architectures.