Source-linked AI summary
DeepRadioID: Real-Time Channel-Resilient Optimization of Deep Learning-based Radio Fingerprinting Algorithms
Francesco Restuccia, Salvatore D'Oro, Amani Al-Shawabka, Mauro Belgiovine, Luca Angioloni, Stratis Ioannidis, Kaushik Chowdhury, Tommaso Melodia
TL;DR
Radio fingerprinting must remain accurate despite changing wireless channels, while deep models are difficult to retrain in real time and adversarial imitation remains insufficiently evaluated. DeepRadioID addresses these issues with transmitter-side FIR optimization and evaluates it across laboratory and large-scale wireless datasets, reporting improved fingerprinting accuracy, reduced adversarial accuracy, and a state-of-the-art improvement.
Problem
Deep-learning radio fingerprinting must handle wireless-channel disruption, limited real-time retrainability, and the unresolved possibility of adversarial fingerprint imitation.
Method
DeepRadioID optimizes a transmitter-side FIR filter using receiver feedback to improve CNN fingerprinting without retraining the underlying model.
Results
DeepRadioID increased fingerprinting accuracy by about 35%, 50%, and 58% across three scenarios, reduced an adversary’s accuracy by about 54%, and achieved 27% improvement over the state of the art on a 100-device dataset.
Takeaways & Limitations
The evaluations support DeepRadioID as a channel- and adversary-resilient optimization system for deep-learning-based radio fingerprinting.
Takeaways & Limitations
The waveform optimization problem is difficult because the CNN objective is highly nonlinear and depends on millions of parameters, while the BER constraint is generally nonlinear and device-specific.
Abstract
from arXiv · showhide
Radio fingerprinting provides a reliable and energy-efficient IoT authentication strategy. By mapping inputs onto a very large feature space, deep learning algorithms can be trained to fingerprint large populations of devices operating under any wireless standard. One of the most crucial challenges in radio fingerprinting is to counteract the action of the wireless channel, which decreases fingerprinting accuracy significantly by disrupting hardware impairments. On the other hand, due to their sheer size, deep learning algorithms are hardly re-trainable in real-time. Another aspect that is yet to be investigated is whether an adversary can successfully impersonate another device fingerprint. To address these key issues, this paper proposes DeepRadioID, a system to optimize the accuracy of deep-learning-based radio fingerprinting algorithms without retraining the underlying deep learning model. We extensively evaluate DeepRadioID on a experimental testbed of 20 nominally-identical software-defined radios, as well as on two datasets made up by 500 ADS-B devices and by 500 WiFi devices provided by the DARPA RFMLS program. Experimental results show that DeepRadioID (i) increases fingerprinting accuracy by about 35%, 50% and 58% on the three scenarios considered; (ii) decreases an adversary's accuracy by about 54% when trying to imitate other device fingerprints by using their filters; (iii) achieves 27% improvement over the state of the art on a 100-device dataset.
1 INTRODUCTION
DeepRadioID addresses IoT authentication needs by using deep-learning radio fingerprinting while targeting channel disruption, real-time optimization, and adversarial imitation. The paper introduces an FIR-based system and evaluates it across 20 software-defined radios and two 500-device datasets.
- Motivation: Deep learning enables general-purpose radio fingerprinting from unprocessed I/Q samples, unlike protocol-specific feature-extraction techniques.The approach targets devices operating across different wireless standards.
- Challenges: Wireless channels can significantly reduce fingerprinting accuracy by disrupting hardware impairments under changing conditions.A five-device experiment compared data collected shortly after training with data collected seven days later.
- Challenges: Real-time retraining is impractical for large deep-learning models, while prior work had not established whether adversaries can imitate legitimate fingerprints.These constraints motivate optimization without retraining and explicit evaluation of adversarial actions.
- Contribution: DeepRadioID uses a carefully optimized transmitter-side digital FIR filter to compensate for current channel conditions without retraining the underlying model.The receiver computes the optimal FIR and sends it back to the transmitter as feedback.
- Evaluation: The evaluation covers 20 nominally identical software-defined radios plus WiFi and ADS-B datasets containing 500 devices each.The 500-device datasets were collected in the wild through DARPA’s RFMLS program.
2 RELATED WORK
Prior radio-fingerprinting work largely relies on protocol-specific physical-layer features and has limited evidence under real-world channel effects. DeepRadioID instead targets general-purpose, scalable, real-time optimization and addresses limitations in earlier impairment-based approaches.
- Existing approaches: Existing fingerprinting methods predominantly extract carefully tailored physical-layer features such as frequency offsets, transients, and protocol-specific signals.Examples include WiFi pilots, training symbols, ZigBee modulation, and WiFi scrambling seeds.
- Evaluation gaps: Some prior evaluations used small testbeds or RF-insulated environments, leaving effectiveness under real-world channel effects less established.One cited study evaluated four ZigBee transmitters, while another considered 130 WiFi cards without channel effects.
- Limitations: Protocol-specific feature extraction limits applicability when IoT devices operate under different wireless standards.The related-work discussion identifies this as a key drawback of feature-based fingerprinting.
- Positioning: DeepRadioID is positioned as a general-purpose and scalable alternative that also addresses real-time accuracy optimization.The related-work discussion contrasts this objective with existing feature-based methods.
- Closest prior work: The closest prior approach improved accuracy through artificial transmitter-side impairments but could not accurately compensate them at the receiver.The paper also identifies an incomplete relationship between hardware impairment and accuracy in that work.
3 DeepRadioID: AN OVERVIEW
DeepRadioID adapts deep-learning fingerprinting to non-stationary wireless channels by optimizing transmitter-side FIR filtering rather than retraining the CNN. The receiver iteratively computes and returns filters, while the design considers adversarial reuse of a legitimate device’s filter.
- Key intuitions: Wireless channels and hardware impairments vary over time, so fingerprinting accuracy requires real-time optimization.Relevant factors include channel conditions, local-oscillator frequency, and RF-circuit temperature.
- Key intuitions: DeepRadioID uses FIR filtering to restore channel-disrupted hardware-impairment patterns that a CNN can recognize.FIR filters are described as easy to implement, with application complexity O(m) for length m.
- Optimization: The optimization seeks FIR taps that maximize the CNN activation associated with a target device while preserving communication performance.The taps are selected by exploiting differentiability through the CNN-based objective.
- System walk-through: The receiver uses CNN outputs and I/Q samples to optimize the target device’s FIR and sends the resulting filter back to the transmitter.The iterative walk-through begins with the previous filter, processes the received signal, and returns the next optimized filter.
- Adversarial setting: An adversary may eavesdrop on a legitimate device’s FIR filter, but the filter is tailored to that device’s channel and impairment conditions.The overview states that reusing the filter does not improve the adversary’s fingerprinting accuracy.
4 DeepRadioID FIR OPTIMIZATION
DeepRadioID uses adaptive FIR filtering to modify transmitted I/Q waveforms, optimizing device-specific CNN activation under channel variation while preserving communication quality. The optimization operates over input slices, uses gradient-based computation, and periodically updates filter taps.
- 4.2 FIR-based Waveform Modification: Channel equalization is insufficient under rapidly changing conditions, so DeepRadioID dynamically adapts waveform modification to compensate for channel effects.The system is designed to address phase and amplitude variations caused by fast fading and multipath.
- 4.2 FIR-based Waveform Modification: DeepRadioID inserts a causal FIR filter before the CNN to modify baseband I/Q samples for real-time adaptive waveform optimization.The filter uses a finite sequence of taps and can manipulate complex signal amplitude and phase.
- 4.3 Waveform Optimization Problem (WOP): The FIR optimization targets higher classification activation for device A while constraining the resulting bit-error rate below BERmax.The objective averages activation over a slice of consecutive inputs, with the BER constraint applied to every input in that slice.
- 4.3 Waveform Optimization Problem (WOP): The WOP is challenging because the CNN objective and BER constraint are highly nonlinear and depend on many model- and device-specific parameters.The CNN may contain millions of parameters, while BER depends on factors including modulation, coding, transmission power, and SNR.
- 4.3.2 Filter taps computation through NCG.: Back-propagation and the chain rule provide gradients with respect to FIR taps, enabling gradient-based solution of the waveform optimization problem.The paper uses these gradients in a Nonlinear Conjugate Gradient algorithm, while allowing other optimization methods in principle.
- 4.3.2 Filter taps computation through NCG.: DeepRadioID periodically updates target-device filter taps when cyclic or occasional triggering events occur.Triggers include timer expiration or target-device accuracy falling below a desired threshold.
5 DeepRadioID FIR COMPENSATION
DeepRadioID uses FIR waveform filtering to optimize fingerprinting while compensating the filter at the receiver, preserving communication quality across modulation and coding schemes. Experiments show limited packet-error and throughput degradation under typical filter magnitudes.
- Motivation: FIR waveform filtering can improve classification but may increase BER by moving I/Q symbols and complicating the nonlinear BER constraint.The resulting optimization must remain generic across modulation, coding, and channel distributions.
- Compensation method: DFT-based receiver compensation reconstructs the original unfiltered I/Q sequence and removes the filter parameter from the BER constraint.The compensation is derived from the received signal’s frequency-domain components.
- Compensation method: FIR compensation is independent of the underlying modulation and coding scheme, making it a general approach for satisfying the BER constraint.The paper presents this property as enabling use across different communication settings.
- Assumptions: Receiver-side compensation relies on estimating the channel and noise frequency responses, which modern wireless networks can usually estimate fairly accurately.Perfect estimation is not assumed.
- Experimental impact: With FIR magnitude parameter ϵ = 0.5, PER increases about 6% and throughput decreases 0.5 kbit/s; typically ϵ < 0.2 limits changes to below 1% and 0.2 kbit/s.The reported typical throughput loss is 0.2%.
6 EXPERIMENTAL RESULTS
DeepRadioID is evaluated on a 20-radio experimental testbed and large WiFi and ADS-B datasets, using live-collected data and multiple accuracy settings. The results show substantial accuracy gains, reduced adversarial imitation, and dataset-dependent convergence behavior.
- 6.3 DeepRadioID Testbed Results: 57% PSA and 35% PBA average improvements were obtained on the experimental testbed.The measurements used live-collected data from nominally identical radios and evaluated short-term and long-term accuracy.
- 6.4 DeepRadioID Dataset Results: On WiFi-100, 10 FIR taps produced about 30% PBA improvement and approximately 74% average accuracy, versus 47% for the state-of-the-art comparison.The model was trained on 100 devices to compare against prior results reported for 93 devices.
- 6.4 DeepRadioID Dataset Results: On ADS-B-500, 10 filter taps produced about 10% PSA and 50% PBA improvement over baseline PSA 0.5028 and PBA 0.6193.The larger PBA gain was attributed to the relationship between small PSA increases and batch accuracy in the 500-device setting.
- 6.4 DeepRadioID Dataset Results: Increasing inputs per slice significantly affected fingerprinting accuracy, especially PBA, across ADSB-500 and WiFi-500.The reported explanation combines more effective FIR averaging with the boosting effect of additional slices.
- 6.4 DeepRadioID Dataset Results: PSA improvement converged near 0.15 for both ADSB-500 and WiFi-500, but WiFi-500 converged faster.ADSB-500 used unprocessed I/Q samples, so its FIR had to compensate for more channel action.
- 6.4 DeepRadioID Dataset Results: The maximum FIR ε-value remained below 0.2, while ADSB-500 required larger average taps than WiFi-100 and WiFi-500.The larger ADSB-500 taps were associated with modifying unprocessed time-domain I/Q samples.
7 CONCLUSIONS
DeepRadioID optimizes deep-learning-based radio fingerprinting accuracy across an experimental testbed and WiFi and ADS-B datasets. It improves fingerprinting performance, reduces adversarial imitation accuracy, and exceeds the state of the art on a 100-device dataset.
- 35%, 50% and 58% increases in fingerprinting accuracy were achieved across the three scenarios considered.
- 54% lower adversary accuracy was achieved when imitating other devices’ fingerprints using their filters.
- 27% improvement over the state of the art was achieved on a 100-device dataset.