Introduction: Toward Real-Time RF Signal Intelligence in a Contested Spectrum
The electromagnetic spectrum has become one of the most contesteddomains in modern warfare, joining air, land, sea, cyber, and space as a theater of strategic competition. The complexity and density of this spectrum have exploded with the proliferation of autonomous platforms, low-power emitters, frequency-agile devices, and non-cooperative actors. The result is an RF environment that is both an operational necessity and a dynamic threat surface.
While traditional spectrum monitoring systems rely on static digitalsignal processing (DSP) pipelines and predefined emitter libraries, they are increasingly inadequate in the face of unknown modulations, frequency hopping, and adversarial signal injection. These methods impose latency, lack adaptability, and fail to meet the speed-of-engagement requirements in environments where milliseconds determine success or mission failure.
Recent advances in embedded deep learning offer a new paradigm. Systemslike RFLearn and DeepRadio, as presented by Restuccia & Melodia (2019) and Soltani et al. (2019), demonstrate that convolutional neural networks (CNNs) trained on raw I/Q data can classify modulation schemes, detect anomalies, and make RF decisions in microseconds, all running natively on low-power FPGAs or
SoCs. These architectures eliminate the reliance on CPU-bound analysis and feature extraction, allowing signal intelligence to move from cloud or host-based systems to the tactical edge.
At the core of this transition is DeepSig’s OmniSIG®, a platformpurpose-built for real-time RF pattern recognition using deep learning inference at the edge. Rather than treating signal recognition as a
mathematical post-processing task, OmniSIG interprets the RF spectrum as a visual, spatiotemporal learning domain, continuously updating its neural models to adapt to adversarial signal environments.
In this article, we explore the architectural transformation from legacysignal recognition systems to embedded neural spectrum intelligence. We detail how platforms like OmniSIG® meet the perational benchmarks laid out in the U.S. Navy's Need 3: real-time classification, anomaly detection, and edge
deployment. We also show how DeepSig’s model hub, training framework, and inference stack are redefining electromagnetic dominance.
By grounding our discussion in practical, hardware-based studies such asRFLearn and DeepRadio, we place OmniSIG in context as a system not only technically innovative but militarily decisive.
References:
Soltani, S. et al. Real-Time and Embedded Deep Learning on FPGA for RF Signal Classification. https://arxiv.org/pdf/1910.05765.pdf
Restuccia, F., Melodia, T. Big Data Goes Small: Real-Time Spectrum-Driven Embedded Wireless Networking Through Deep Learning in the RF Loop. https://ece.northeastern.edu/wineslab/papers/Restuccia19INFOCOM.pdf
The electromagnetic spectrum is no longer an auxiliary domain-it is the battlespace. In today’s contested multi-domain environment, control over RF signaling is as vital as control over air, land, or sea. From GPS spoofing over the Black Sea to burst transmissions from frequency-hopping UAVs in the
Pacific, RF activity is now a primary tool of both state and non-state actors. These signals carry commands, disrupt navigation, hide drones, disable radar, and enable information warfare in ways that traditional systems were never designed to counter.
Modernadversaries now utilize complex RF techniques such as low-probability-of-intercept (LPI) waveforms, time-varying digital modulations (e.g., adaptive QAM and spread-spectrum BPSK), and encrypted command channels that dynamically adjust carrier frequencies and coding rates. These signals are often transmitted in microbursts and occupy wideband channels with agile center
frequency shifts, making real-time demodulation and attribution extremely difficult. Furthermore, transmission environments are complicated by multipath fading, Doppler shift, co-channel interference, and adversarial signal injection-all of which require intelligent processing beyond classical DSP thresholds.
The U.S.Navy’s Need 3, as articulated by NSWC Dahlgren, outlines a critical requirement: develop machine-learning-based RF monitoring systems that operate in real time, classify signals at the edge, and detect anomalies autonomously under degraded or contested conditions. These systems must be able to classify, cluster, and predict signal behavior on minimal hardware, in disconnected
environments, under rapidly changing operational constraints.
Thisrequirement translates into the deployment of spectrum classifiers that must operate at sub-20 ms inference latencies, execute on edge-optimized hardware such as NVIDIA Jetson, Xilinx Zynq UltraScale+, or ARM Cortex-based SoCs, and be able to distinguish and adapt to modulations including BPSK, QPSK, GFSK, CPM, 16-QAM, and GMSK in environments with SNRs below 0 dB. The Navy’s operational profile further demands support for continual learning-field-deployable
retraining using few-shot datasets captured from novel emitters.
This iswhere DeepSig, a defense-focused AI company based in Arlington, Virginia, steps in. Their flagship product, OmniSIG®, is not just a spectrum analyzer-it is a neural perception engine designed to listen, learn, and make decisions in the invisible domain of RF warfare. It fundamentally replaces legacy DSP-centric architectures with deep neural networks trained to interpret RF as a visual
pattern, enabling advanced classification, anomaly detection, and adaptive learning.
At atechnical level, OmniSIG ingests raw in-phase and quadrature (I/Q) samples from the RF front end, converts them to spectrograms using short-time Fourier transforms (STFT) or constant-Q transforms (CQT), and applies a multi-layer convolutional neural network to classify known modulations and flag unknown or anomalous signals. The model architecture, inspired by work such as O’Shea et al. (2016), typically comprises multiple convolutional layers with ReLU activation, batch normalization, and max pooling, followed by dense output layers for classification across signal types. The trained models are
compressed via quantization and pruning, enabling deployment with <100 MB memory footprint and power usage under 5W on embedded FPGA hardware.
OmniSIG hasdemonstrated classification accuracy above 90% across 20+ modulation schemes in
SNR conditions ranging from -4 dB to +18 dB, with inference latency as low as 8 ms on embedded platforms and throughput exceeding 1 Msamples/sec. Its models are retrainable on-site using the OmniSIG Studio toolkit, which supports labeling, noise modeling, adversarial training, and federated learning without requiring centralized data transfer.
By combining signal understanding with AI-native architectures, OmniSIG transforms spectrum monitoring from reactive filtering into proactive perception, enabling electromagnetic superiority in future conflict domains.
The Limitsof Traditional DSP-Based Signal Intelligence
Legacyelectronic warfare and spectrum analysis systems have long relied on deterministic digital signal processing (DSP) chains to identify known threats. These systems work by executing a fixed pipeline of manual feature extraction-typically involving Fast Fourier Transforms (FFT), cyclostationary analysis, autocorrelation, and matched filtering-followed by a set of rule-based or
heuristic classifiers. This architecture assumes that incoming signals conform to known modulation patterns, are separated in frequency or time, and arrive in predictable forms amenable to template-based detection.
However,the modern electromagnetic environment fundamentally violates these assumptions. RF landscapes are now characterized by extreme density, signal variability, and adversarial manipulation. The static structure of DSP chains is inherently fragile when confronted with evolving RF tactics that
intentionally disrupt or camouflage signal characteristics. Theselimitations manifest in several specific ways:
Modulation Agility – In many contested environments, emitters now dynamically switch between modulation types (e.g., shifting from BPSK to 16QAM or from OFDM to CPM mid-frame) in order to evade detection or fingerprinting. Traditional DSP relies on fixed thresholds and templates which fail to track such mid-transmission shifts, especially when they occur below 10 ms windows.
- Frequency Hopping Spread Spectrum (FHSS) – FHSS systems change carrier frequencies rapidly and pseudorandomly across wide bandwidths, sometimes in sub-millisecond intervals. This breaks the fixed time-frequency windows of FFT-based methods and defeats any attempt at coherent time-series analysis unless the hopping sequence is known in advance-which is rarely the case in adversarial scenarios.
- Intentional Overlapping Signals – Adversaries can flood the RF spectrum with synthetic or decoy signals using multi-tone jamming, chirped interference, or overlapping modulated bursts that mimic friendly systems. The result is that spectrum becomes indistinguishable using logic-based classifiers, which cannot resolve signal boundaries or confidently attribute modulation types when signals collide in time and frequency.
- Adversarial Signal Injection – More recently, deep learning-based adversarial attacks have been
demonstrated in the RF domain. For example, emitters may apply imperceptible perturbations to their waveform (e.g., phase-shift patterns or amplitude masks) that confuse conventional classifiers while remaining demodulable by intended receivers. These perturbations are specifically designed to trigger misclassification in template-based or shallow-learning systems. - Compoundingthese technical challenges is the reliance on static emitter libraries. Traditional EW systems depend on predefined databases of known signal types-often with hard-coded decision trees or lookup tables. This makes them functionally blind to emergent or unknown waveforms. In operational environments such as grey-zone conflict or counterinsurgency operations, where new or unregistered devices are deployed without formal spectrum governance, this rigidity creates unacceptable latency in detection and identification.
Furthermore,these systems lack temporal contextual awareness. A DSP chain may identify a 16QAM waveform, but it cannot determine whether the same emitter previously used QPSK, nor whether the signal pattern suggests reconnaissance, command-and-control, or spoofing behavior. Without stateful inference or sequence modeling, DSP-based intelligence offers only momentary, decontextualized snapshots of the electromagnetic environment.
Finally,the compute burden of traditional DSP pipelines scales poorly in dense environments. As signal density increases, FFT computation and filter banks become resource-bound, leading to latency, dropped packets, or forced downsampling-each of which erodes real-time situational awareness.
In wartimeconditions-where ambiguous, deceptive, and novel signals are the norm-these architectural weaknesses form operational blind spots. The inability to identify unknowns, adapt to dynamic modulation, or infer signal intent degrades the commander's decision-making loop and creates windows of vulnerability that can be exploited by adversaries. Traditional DSP, while still useful for
baseline signal integrity and communications engineering, cannot meet the agility, adaptability, and autonomy required for spectrum dominance in contested battlespaces.
OmniSIG®: ANeural Architecture for the RF Domain
DeepSig’sOmniSIG® platform fundamentally redefines how RF signals are analyzed-not as
deterministic waveforms to be parsed by signal engineering heuristics, but as high-dimensional perceptual structures to be interpreted through machine learning. This transformation shifts RF analysis from rule-based identification toward adaptive pattern recognition, enabling systems to detect, classify, and understand signals in real time with resilience to noise, deception, and unknown protocols.
At itscore, OmniSIG ingests raw in-phase (I) and quadrature (Q) baseband samples-directly from software-defined radios (SDRs), embedded front ends, or hardware capture devices-and converts these temporal sequences into time-frequency spectrograms using short-time Fourier transforms (STFT) or more advanced adaptive methods like constant-Q transform (CQT). These spectrograms are treated as image-like data structures, allowing the application of convolutional neural networks
(CNNs) originally developed for visual classification.
Unliketraditional classifiers, OmniSIG’s neural models learn complex representations that are invariant to time shift, frequency offset, and SNR variability. The result is a platform that doesn’t just recognize modulation types-it learns the signature behavior of emitters across variable channel conditions, antenna patterns, transmission bursts, and intentional obfuscation.
Expanded Core Capabilities
Real-time signal classification
OmniSIG delivers classification results within microsecond to low-millisecond inference latency, depending on hardware. Models have been benchmarked at >90% accuracy across 20+ modulation types (e.g., BPSK, QPSK, GFSK, 16QAM, CPM, OFDM) in SNR ranges from -4 dB to +18 dB. This allows fast, confident labeling of signals in motion-critical for cueing countermeasures, filtering threat traffic, or triggering alerts.
Anomaly detection
OmniSIG embeds unsupervised clustering and outlier detection mechanisms directly in the spectral latent space. These functions monitor the statistical distribution of activation patterns in the neural layers, flagging any signals that fall outside known modulation classes or deviate from emitter-specific norms. This is vital for catching novel jamming techniques, spoofed control links, or unexpected emissions from compromised assets.
Signal fingerprinting
The platform constructs fine-grained embeddings of RF emissions that reflect not only modulation, but also transient hardware-specific traits such as power amplifier drift, clock jitter, and phase noise. Thesefeatures enable RF fingerprinting of emitters, allowing attribution across missions or geographic regions-even under intentional masking or low-SNR conditions.
Efficient edge inference
OmniSIG models are optimized through quantization-aware training, pruning, and hardware-specific compilation (e.g., Xilinx Vitis AI, Nvidia TensorRT) to run efficiently on edge platforms like Jetson Nano, Xilinx Zynq UltraScale+, and ARM Cortex-A embedded modules. Typical power consumption remains under 5W for full-spectrum classification pipelines, supporting continuous operation on tactical UAVs, man-portable systems, or maritime ISR units.
On-device retraining and domain adaptation
Using the OmniSIG Studio toolkit, forward-deployed operators can label new I/Q samples, augment training sets with noise injection or adversarial perturbations, and retrain models in-theater without backhaul to cloud or base servers. This supports continual learning and rapid adaptation to new emitter types, mission profiles, or adversarial tactics.
Architectureand Training Pipeline
OmniSIG’smodel architecture is deeply informed by developments in computer vision, specifically architectures designed for robustness under distortion, occlusion, and spatial deformation. Spectrogram input tensors typically take the form of 2D or 3D structures, with input channels representing amplitude, phase, or signal energy features. The core CNN stack may include:
Multiple convolutional layers with ReLU or Leaky ReLU activation
- Batch normalization for regularization and faster convergence
- Max pooling for translation invariance
- Fully connected layers for feature integration
- Training datasets are composed of both synthesification or embedding learning
Training datasets are composed of both synthetically generated signals (via GNU Radio,
MATLAB, or domain-specific RF modeling engines) and real-world emissions captured from over-the-air sources using SDR platforms like USRP, HackRF, and commercial radios. These datasets are heavily augmented with additive white Gaussian noise (AWGN), multipath fading models, Doppler shifts, frequency offsets, and time-domain jitter to ensure generalization.
DeepSig’smodels are further enhanced using techniques such as:
Knowledge distillation for compact student models
- Adversarial training to defend against spoofing and signal perturbation
- Federated learning for distributed model updates without centralized data transfer
- This hybridtraining strategy ensures that OmniSIG’s inference models maintain operational
robustness, even in contested and deceptive signal environments where the RF landscape is dynamic and intentionally manipulated.
By fusingvisual feature extraction, temporal dynamics, and domain-specific augmentations, OmniSIG builds a complete pipeline for RF signal intelligence-from waveform ingestion to signal classification and emitter behavior modeling-entirely within the AI-native paradigm.
Technical Innovation: Spectrum Intelligence at Machine Speed
What setsOmniSIG apart is its foundational reimagining of RF analysis as a machine perception task, where radio signals are treated not as numeric waveforms to be demodulated, but as high-dimensional visual patterns to be learned. This paradigm shift allows OmniSIG to perform classification, anomaly detection, and pattern recognition in signal environments that overwhelm traditional systems. Instead of relying on handcrafted feature pipelines, OmniSIG uses end-to-end learned representations that are resilient to interference, distortion, and adversarial manipulation.
Forexample, a BPSK-modulated emitter exposed to broadband jamming will exhibit time-frequency features that appear diffused, distorted, and non-stationary when visualized as spectrograms. To a human analyst or rule-based DSP chain, this may resemble background noise. But to a convolutional neural network trained on diverse spectrogram transformations, the spectral signature of the signal-including its symmetry, power envelope, and spectral roll-off-remains intelligible. OmniSIG’s model identifies the underlying modulation through learned spatial invariants and channel noise robustness, executing this inference in microseconds on embedded silicon.
Demonstrated Performance and Latency
In benchmark scenarios, OmniSIG has consistently achieved:
- Greater than 90% classification accuracy across a modulation suite that includes BPSK, QPSK, GMSK, GFSK, 16QAM, CPM, and OFDM
- Stable performance across wide SNR ranges, from as low as -4 dB to as high as +18 dB, under additive white Gaussian noise (AWGN), Rayleigh fading, and simulated adversarial
interference. - Inference latencies below 20 milliseconds on low-power devices such as the NVIDIA Jetson Nano and Xilinx ZCU104 boards, enabling real-time operation on tactical edge nodes
with <10W power budgets. - This performance profile allows OmniSIG to be deployed in environments with tight
response requirements-such as autonomous ISR assets, dismounted EW operators,
or shipboard SIGINT platforms-without sacrificing computational feasibility or
power efficiency.
Training at Scale: Model Fidelity and Generalization
The modeltraining process behind OmniSIG leverages hundreds of gigabytes of synthetic
I/Q datasets, generated across multiple sampling rates, modulation types, and channel conditions. These datasets are built using tools such as GNU Radio, MATLAB Communications Toolbox, and custom RF emulation environments that simulate jamming, burst transmission, frequency drift, pulse shaping, and multi-path effects. Each dataset is augmented with:
- Randomized channel models (e.g., Rician, Rayleigh, Nakagami)
- Adversarial overlays (e.g., spectral masking, noise injection, phase warping)
- Time-frequency shifting, jitter, and amplitude scaling
- To ensureoperational relevance, DeepSig also incorporates real-world over-the-air
captures collected using SDRs from urban, maritime, and suburban environments. These emissions are labeled, clustered, and fused into the training regime to support domain adaptation and real-signal generalization.
OmniSIGStudio: Integrated Mission-Focused ML Tools
OmniSIG’svalue extends beyond classification performance. With OmniSIG Studio, engineers
and operators are equipped with a full development pipeline tailored for spectrum mission engineering. This includes:
- Interactive environment simulation: Inject, manipulate, and visualize synthetic and real-world
signal environments under various SNR and interference levels - Labeling and annotation tools: Create supervised datasets from field-captured emissions
- Training, validation, and model evaluation: Run domain-specific training cycles with performance
monitoring, confusion matrices, and misclassification analysis - Export-ready model compilation: Output edge-ready neural models in formats optimized for TensorRT, ONNX, or Xilinx DPU backends
- Theplatform also supports real-time anomaly detector deployment, enabling
persistent monitoring for new or out-of-distribution signals. Models can be
pushed to hardware platforms through secure containers or manually updated in
disconnected environments via physical transfer.
OmniSIGModel Hub: Pretrained Operational Models
Toaccelerate time-to-deployment, DeepSig provides the OmniSIG Model Hub-a library of pretrained neural models for various mission types, including maritime surveillance, urban SIGINT, ISR for unmanned systems, and base defense. These models are pre-optimized for different computer platforms and cover modulation profiles commonly seen in military, commercial, and adversarial communications.
Bycombining inference-optimized models, scalable training pipelines, embedded toolchains, and domain-specific customization, OmniSIG transforms RF monitoring into an AI-native capability that meets the operational latency, fidelity, and adaptability demands of future warfighting. It is not just spectrum sensing-it is spectrum cognition.
Deployment Scenarios
OmniSIG hasbeen purpose-built for highly distributed, mission-critical deployments where environmental complexity, compute constraints, and adversarial pressure converge. Its low-SWaP (Size, Weight, and Power) neural inference architecture and embedded retraining capabilities make it a highly adaptive solution for near-peer conflict, gray-zone operations, and autonomous platform integration. It is not just deployable-it is operationally mobile, modular, and mission-adaptable.
MaritimeSurveillance and Naval ISR
Onboard maritimevessels, OmniSIG provides persistent electromagnetic overwatch for detecting
and classifying RF signals associated with UAV control links, spoofed AIS beacons, SATCOM interference, or hostile surface radar pings. In littoral zones or contested sea lanes, these signals often emerge briefly and in cluttered spectral conditions. OmniSIG enables vessels to:
Identify low-SNR UAV telemetry bursts hopping across bands like 900 MHz or 2.4 GHz
- Detect AIS spoofing patterns via anomalous modulation or burst repetition rates
- Monitor non-cooperative emitters, including pirate or smuggler comms operating under commercial cover
- Its edgeinference models allow classification and fingerprinting to occur in real time,
even on platforms with intermittent network links to central command.
Ground-BasedTactical Surveillance
In forward-deployedunmanned ground systems, such as perimeter security bots or vehicle-mounted ISR units, OmniSIG enables localized detection of RF anomalies linked to IED triggers, adversary relay nodes, or clandestine command networks. These platforms often operate in unstructured terrain where bandwidth to central processing is minimal or absent.
OmniSIG’s capacity to:
Detect short-duration modulated pulses used in RF-based IED detonators
Monitor for encrypted handset bursts in ultra-high frequency (UHF) ranges
- Execute classification on embedded FPGAs with latency under 10 ms makes it a decisive capability for autonomous security operations in austere or denied environments.
Space-BasedISR and Spectrum Mapping
Onboard Low-EarthOrbit (LEO) satellites, OmniSIG supports wideband RF mapping and signal
classification across global regions. Space-based platforms provide unique electromagnetic vantage points-but are constrained by:
- Power budgets under 20W
- Limited onboard storage
- High latency or intermittent downlink windows
OmniSIGallows onboard inference-level analysis of spectrum data collected across
commercial, military, and unregistered emitters. This includes:
- Mapping dynamic RF occupancy patterns in urban warzones
- Detecting emitter migration as forces maneuver
- Identifying burst-mode tactical datalinks used by high-speed aircraft or missiles
- The abilityto flag signals in real time and transmit only prioritized metadata back to
Earth drastically reduces bandwidth demands and increases targeting timelines.
SpecialOperations and Denied-Area Reconnaissance
For SpecialOperations Forces (SOF) and intelligence teams operating in communications-denied or GPS-degraded environments, OmniSIG provides a lightweight RF reconnaissance tool that enables:
- Real-time identification of hostile radios operating under cover protocols Detection of jammer activation signatures
- On-device retraining to new emissions without network access
- Ruggedizededge nodes (e.g., Jetson Xavier NX, Xilinx RFSoCs) can be integrated into tactical packs, vehicles, or drone payloads, with mission-specific models deployed via pre-mission provisioning or offline USB load.
StrategicProgram Relevance
Thesedeployment configurations align OmniSIG directly with:
JADC2 (Joint All-Domain Command and Control) initiatives requiring multi-domain situational awareness and cross-platform signal intelligence
- Electromagnetic Maneuver Warfare (EMW) doctrines focusing on agile, resilient, and informed
spectrum operations - AI-enabled Electronic Warfare (AI-EW) automation efforts tasked with reducing operator burden and shortening the sensor-to-decision timeline
- OmniSIG’sability to operate under constrained power, process in real time, adapt to new
threats, and integrate with autonomous platforms makes it a keystone technology for the evolving concept of cognitive EW-where systems don’t just see the spectrum, they understand it.
Direction:Towards Autonomous RF Cognition
As the character of warfare continues to evolve toward multi-domain, software-defined, and electronic-first battlefields, the demand for machine-perceptive RF ntelligence has become operationally non-negotiable. The electromagnetic spectrum is no longer merely a medium for communication-it is an activelycontested and adversarial domain where delay, misclassification, or signal blindness can result in compromised missions or lost platforms. In this space,human speed is no longer sufficient. Manual workflows, static emitterlibraries, and rules-based detection are structurally incapable of matching theagility of peer adversaries who now employ autonomous, encrypted, and frequency-agile RF assets.
DeepSig’s OmniSIG®platform directly responds to this doctrinal shift. Rather than upgradinglegacy DSP workflows, it introduces a fundamentally different approach-one inwhich RF is interpreted by deep neural architectures trained to understand, not decode, electromagnetic activity. These systems ingest raw I/Q samples, translate them into learned spectrogram embeddings, and produce fast,
actionable classifications and anomaly detections with sub-20 millisecond latency. OmniSIG is more than a spectrum analyzer-it is a real-time, edge-native cognition engine purpose-built for decision dominance in the RF domain.
Its key strength lies in the fusion of tactical deployability with cognitive adaptability. The platform supports continuous on-device retraining, federated learning, adversarial resilience, and fine-grained signal fingerprinting, all within a compact model architecture that runs on low-power SoCs or eembedded GPUs. With OmniSIG Studio and Model Hub, warfighters gain a full ecosystem of training tools, pre-optimized models, and deployment support-enabling rapid field adaptation without dependency on centralized infrastructure.
In future electromagnetic operations, success will belong not to the side with the loudest signal, but to the side with the fastest and most intelligent interpretation of the spectrum. OmniSIG enables platforms to perceive patterns in crowded spectral space, adapt to unseen threats, and respond before human operators could even isolate the waveform.
It is notsimply a tool for signal recognition-it is the beginning of autonomous RF cognition: a new operational paradigm where machines not only see the invisible domain, but understand and dominate it in real time. This is the next frontier in electromagnetic warfare, and OmniSIG is already operating on that edge.
* * *
References:
- DeepSig. OmniSIG. 2025. https://www.deepsig.ai/omnisig/
- DeepSig. AI Native Wireless Communications. 2025. https://www.deepsig.ai/
- DeepSig. OmniSIG Model Hub. 2024. https://www.deepsig.ai/omnisig-model-hub-access-industry-leading-ai-ml-spectrum-sensing-models/
- DeepSig. OmniSIG Studio. 2025. https://www.deepsig.ai/omnisig-studio/
- DeepSig. Transforming Data into Actionable Intelligence. 2024. https://www.deepsig.ai/transforming-data-into-actionable-intelligence/
- DeepSig. OmniSIG DF. 2025. https://mclct.ro/wp-content/uploads/2025/03/OmniSIG-DF.pdf
- Bertozzi A. et al. LibIQ: Toward Real Time Spectrum Classification in O RAN dApps. 2025. https://arxiv.org/html/2505.10537v1
- Shi L. et al. A Survey on State of the Art Deep Learning Applications. 2024. https://arxiv.org/html/2403.17561v7
O’Shea T. J. et al. Convolutional RadioModulation Recognition Networks. 2016. https://arxiv.org/abs/1602.04105