Paper deep dive
The S-ICDF Dataset: Sionna-Simulated Dynamic Interference Characterization and Direction Finding
Christian Wielenberg, Lucas Heublein, Jonathan Ott, Alexander Mattick, Nisha L. Raichur, Jonas Pirkl, Lukas Schelenz, Tobias Feigl, George Yammine, Christopher Mutschler, Felix Ott
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 95%
Last extracted: 7/7/2026, 7:53:54 AM
Summary
The paper introduces S-ICDF, a large-scale simulated indoor interference dataset generated using the Sionna library to address data scarcity in GNSS jamming and spoofing research. It covers 102 interference configurations with diverse parameters and benchmarks classical signal processing and machine learning methods for interference characterization and direction finding.
Entities (8)
Relation Signals (9)
S-ICDF Dataset → generatedby → Sionna
confidence 97% · S-ICDF, a large-scale indoor interference dataset generated with Sionna, a GPU-accelerated simulation library
S-ICDF Dataset → covers → 102 interference configurations
confidence 96% · S-ICDF covers 102 interference configurations, including diverse antenna array patterns, bandwidths, and simulation settings
S-ICDF Dataset → benchmarkedwith → ESPRIT
confidence 95% · We further provide baseline results by benchmarking S-ICDF with classical estimation and direction finding (DF) methods (MUSIC, ESPRIT, and CAPON)
S-ICDF Dataset → benchmarkedwith → MUSIC
confidence 95% · We further provide baseline results by benchmarking S-ICDF with classical estimation and direction finding (DF) methods (MUSIC, ESPRIT, and CAPON)
S-ICDF Dataset → benchmarkedwith → CAPON
confidence 95% · We further provide baseline results by benchmarking S-ICDF with classical estimation and direction finding (DF) methods (MUSIC, ESPRIT, and CAPON)
S-ICDF Dataset → benchmarkedwith → XceptionTime
confidence 94% · We evaluate interference characterization and DF performance using both classical signal processing baselines and ML-based approaches. ... we employ an XceptionTime architecture
S-ICDF Dataset → usedfor → Direction Finding
confidence 94% · The S-ICDF dataset is released as a public benchmark to facilitate reproducible evaluation and comparison in future research, offering broad and controlled coverage of interference conditions, array-based IQ measurements, and realistic time-varying multipath
S-ICDF Dataset → usedfor → Interference Characterization
confidence 94% · We evaluate interference characterization and DF performance using both classical signal processing baselines and ML-based approaches.
Jamming → threatens → GNSS
confidence 93% · Jamming devices interfere with global navigation satellite system (GNSS) signals emitted by RF sources and represent a serious threat
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Jamming and spoofing threaten wireless and satellite navigation by disrupting or manipulating radio frequency (RF) signals, undermining availability, integrity, and trust. Robust interference monitoring (i.e., detection, classification, characterization, and direction finding) is therefore essential to identify and localize anomalous signals. While machine learning (ML) promises improved performance in complex environments, its development and validation depend on large-scale datasets that capture realistic signal and channel variability. Collecting such data in the real world is difficult because intentional jamming is illegal and ground-truth attribution is confounded by propagation, hardware, and environmental effects. To address this gap, we create and publish S-ICDF, a large-scale indoor interference dataset generated with Sionna, a GPU-accelerated simulation library for physical-layer wireless communications. S-ICDF covers 102 interference configurations, including diverse antenna array patterns, bandwidths, and simulation settings such as noise level and reflection depth. We further provide baseline results by benchmarking S-ICDF with classical estimation and direction finding (DF) methods (MUSIC, ESPRIT, and CAPON) and with modern ML approaches. The dataset is publicly available at: this https URL
Tags
Links
- Source: https://arxiv.org/abs/2607.03411v1
- Canonical: https://arxiv.org/abs/2607.03411v1
Trouble viewing inline? Open PDF directly →
Full Text
36,635 characters extracted from source content.
Expand or collapse full text
The S-ICDF Dataset: Sionna-Simulated Dynamic Interference Characterization and Direction Finding Christian Wielenberg, Lucas Heublein, Jonathan Ott, Alexander Mattick, Nisha L. Raichur, Jonas Pirkl, Lukas Schelenz, Tobias Feigl, George Yammine, Christopher Mutschler, Felix Ott Fraunhofer Institute for Integrated Circuits IIS, 90411 N ̈ urnberg, Germany christian.wielenberg, tobias.feigl, george.yammine, christopher.mutschler, felix.ott@iis.fraunhofer.de Abstract—Jamming and spoofing threaten wireless and satellite navigation by disrupting or manipulating radio frequency (RF) signals, undermining availability, integrity, and trust. Robust interference monitoring (i.e., detection, classification, character- ization, and direction finding) is therefore essential to identify and localize anomalous signals. While machine learning (ML) promises improved performance in complex environments, its development and validation depend on large-scale datasets that capture realistic signal and channel variability. Collecting such data in the real world is difficult because intentional jamming is illegal and ground-truth attribution is confounded by propaga- tion, hardware, and environmental effects. To address this gap, we create and publish S-ICDF, a large-scale indoor interference dataset generated with Sionna, a GPU-accelerated simulation li- brary for physical-layer wireless communications. S-ICDF covers 102 interference configurations, including diverse antenna array patterns, bandwidths, and simulation settings such as noise level and reflection depth. We further provide baseline results by benchmarking S-ICDF with classical estimation and direction finding (DF) methods (MUSIC, ESPRIT, and CAPON) and with modern ML approaches. The dataset is publicly available at: https://gitlab.c-asp.fraunhofer.de/darcy gnss/sicdfdataset Index Terms—Sionna, Simulation, Interference Monitoring, Characterization, Direction Finding, Localization, Dataset, IQ I. INTRODUCTION Jamming devices interfere with global navigation satellite system (GNSS) signals emitted by RF sources and represent a serious threat, undermining the robustness and reliability of both precise positioning and wireless communications [1]– [3]. This problem has grown severe in recent years as readily obtainable jammers have become increasingly common [4], [5]. As a result, interference detection, classification [6], characterization [7], localization [8], [9], and mitigation [10] have emerged as key research areas. Numerous solutions have been explored, spanning traditional signal processing techniques [11]–[13] as well as ML-based methods [14], [15]. However, training such ML-based methods (particularly with the emergence of large foundation models) demands large-scale datasets that not only match the target task, but also comprehensively cover and remain well-balanced across all parameters the model is intended to learn, ensuring that relevant conditions and variations are adequately represented rather than dominated by a small subset of frequent cases [16]. In the context of interference monitoring, this entails broad and uniform coverage of jammer characteristics, including waveform/type (e.g., continuous wave, noise, swept, pulsed), center frequency and bandwidth, temporal behavior (duty cycle, burst/sweep dynamics), transmit power or interference- to-noise-ratio (INR), and spatial/propagation conditions (e.g., location and mobility, (non-)line-of-sight (LoS/NLoS), and multipath [7]. Collecting real-world recordings involving jam- ming is often legally restricted [17], and typically necessitates extensive measurement campaigns, such as the Jammertest in Norway [18]. Hence, our objective is to generate a simulated dataset and to benchmark ML models on this dataset to enable rapid model iteration and development. Although publicly available (simulated) datasets exist [19]–[25], they typically exhibit at least one of the following limitations: (1) insufficient coverage of a broad range of interference characteristics; (2) the absence of multi-patch antenna measurements required by DF methods; or (3) stationary receiver setups that preclude time-series-based DF approaches. To overcome these shortcomings, we employ Sionna [26] as the underlying simulation environment. Fig. 1: Sionna simulation with ray tracing. Contributions. In the fol- lowing, we summarize our main contributions: (1) In Sionna, we model an indoor industrial environment and employ ray tracing to sim- ulate propagation from an interference source and the resulting signals received by an antenna array (refer to Figure 1). (2) We consider 110 distinct interference pa- rameterizations, targeting both interference characterization and direction-of-arrival (DoA) estimation between the source and the receiver. The receiver is modeled as a 2×2-multi-patch array providing raw in-phase and quadrature (IQ) samples. (3) To support time-series-based DF via a synthetic aperture effect, the antenna platform moves through the environment during data acquisition. (4) We evaluate interference char- acterization and DF performance using both classical sig- nal processing baselines and ML-based approaches. The S- ICDF dataset is released as a public benchmark to facilitate reproducible evaluation and comparison in future research, offering broad and controlled coverage of interference condi- tions, array-based IQ measurements, and realistic time-varying multipath due to ray-traced propagation and receiver motion. arXiv:2607.03411v1 [eess.SP] 3 Jul 2026 TABLE I: Overview of publicly available real-world and simulated GNSS interference datasets. DatasetReal-world vs. SimulatedInterference Type Data ModalityDirection Finding Suitability TEXBAT [19]Real-world (field recordings) SpoofingIQ recordingsPrimarily single-channel Oak Ridge (OAKBAT) [25] Real-world (digitized)SpoofingDigitized RF/IQ recordingsLimited (mainly single-channel) L.I.N.K. Hall [9]Real-world (indoor)Moving jammerIQ recordings2×2-array for DF, not phase-coherent Tuni2025 (TG-GGSD) [21] Real-world (lab)SpoofingIQ recordings + scenario metadata Single stream; not DF-centric In-lab DF Validation [22]Real-world (lab)JammingMeasurement-based evaluationModerate (DF evaluated) GATEMAN [23]Real-world (in-lab)JammingWideband IQ recordingsLimited Evil WaveForms [24]Simulated (GNSS simulator) Intf. waveformsSimulated IQ waveformsLimited (no array/motion labels for DF) Raw IQ Dataset [20]SimulatedJammingLabeled IQ samplesNo array/multi-sensor information I. RELATED WORK Table I provides an overview of publicly available datasets for GNSS interference monitoring differing substantially in how threats are generated and what they enable: TEX- BAT [19], OAKBAT [25], and Tuni2025 [21] provide scenario- based RF/IQ recordings primarily targeting spoofing analysis and benchmarking, whereas the IQ jamming classification dataset [20] offers labeled samples geared toward supervised jammer-type recognition. For jamming-focused studies, the GATEMAN release [23] supplies wideband in-lab recordings of GNSS and jammer signals, complemented by the exper- imental validation study in [22] that evaluates detection and DF-methods under controlled conditions. The Evil WaveForms work [24] is simulator-based and provides synthetic threat waveforms for reproducible testing, but lacks the array and motion information required for robust DF/localization bench- marking. Heublein et al. [9] presented a large-scale dataset of moving GNSS jamming devices recorded in an indoor in- dustrial environment with dynamic multipath, providing multi- patch raw IQ measurements together with ground-truth relative position labels for jammer DF and localization. While useful as benchmarks, most datasets either lack the synchronized multi-antenna phase information or the trajec- tory/pose ground truth needed to train and evaluate DF under moving antennas or jammers. Moreover, motion diversity and key dynamic effects (time-varying multipath, LoS/NLoS changes, and broad jammer parameter/INR coverage) are often limited, restricting robust generalization and detailed compar- isons. To address these limitations and to enable controlled, scalable generation of motion-rich, fully labeled array mea- surements under realistic propagation, we simulate our dataset using Sionna [26]. Prior work has primarily used Sionna for large-scale synthetic channel and radio-map generation in digital-twin settings, whereas we are, to the best of our knowledge, the first to leverage Sionna to generate time-series raw IQ measurements for DF with a moving antenna array in the presence of interference sources. Research on GNSS interference monitoring is advanced, with a wide range of well-established signal processing and ML-based techniques. Accordingly, we utilize well-established baselines to ensure comparable evaluation: for characterization tasks, we employ an XceptionTime [27] architecture, while for DF we consider classical estimators, including MUSIC [28], ESPRIT [29], and CAPON [30], which are commonly used reference methods for DoA estimation. I. SIMULATION In this section, we present the complete simulation-to- benchmark workflow. We first describe the end-to-end data- generation pipeline (Sec. I-A). We then detail the Sionna- based simulation setup, including the ray-tracing environment and all relevant configuration parameters (Sec. I-B). Finally, we introduce the S-ICDF dataset and summarize its structure, parameter ranges, and data splits (Sec. I-C). A. Pipeline The simulation pipeline is designed to systematically quan- tify how individual simulation parameters affect both the received signal structure and the performance of direction find- ing algorithms. Figure 2 provides an overview. First, a param- eter configuration is specified by defining the range and reso- lution of each variable of interest. To enable an interpretable sensitivity analysis, we vary only a single parameter at a time while holding all remaining parameters fixed at their default values, such that any performance variation can be attributed to that parameter. Each configuration is then provided to Sionna, which models the transmission environment (including propagation and channel effects) and generates corresponding complex baseband signals (see Section I-B). From the sim- ulator output, we extract raw IQ samples, preserving the full complex-valued information required by downstream localiza- tion methods (see Section I-C). The resulting IQ sequences are subsequently processed by the algorithms under study, including classical baseline algorithms and the considered ML models trained directly on the simulated data (see Section IV). Finally, performance metrics are computed and aggregated across parameter settings, enabling a systematic comparison of the impact of each parameter on MUSIC, ESPRIT, CAPON and ML-based localization performance. B. Sionna Simulation Environment NVIDIA Sionna [26] is an open-source, GPU-accelerated (TensorFlow-based) simulation library for wireless PHY/link- level research where transmitters, channels, and receivers are built as modular, differentiable blocks so the user can run fast Monte-Carlo evaluations and also backpropagate gradi- ents end-to-end for learning-based designs. Sionna includes detailed components for modern chains (e.g., OFDM/MIMO, coding, estimation/detection) and an optional ray-tracing mod- ule (Sionna RT) to generate geometry/material-aware channels from 3D scenes. Parameter Definition Definition of the parameter set Sionna Simulation Ray-tracing & channel generation IQ Extraction Raw baseband data MUSIC / ML Training Feature engineering & optimization Results Characterization / Direction finding / Visualization Fig. 2: Pipeline from parameter definition and Sionna simulation to IQ extraction, MUSIC/ML training, and evaluation. The simulation environment represents an industrial hall- like scenario and is designed to capture the dense multipath conditions typical of indoor industrial settings; an example realization is shown in Figure 1. The scene is implemented using Sionna’s ray-tracing engine, which computes physically grounded propagation paths between transmitter and receiver. It comprises a large rectangular layout with predominantly metallic boundary surfaces (walls, ceiling, and floor) and includes interior structures such as shelving units that serve as reflectors and scatterers. This geometry yields a large number of multipath components, providing a challenging yet realistic testbed for localization methods. The maximum number of reflections considered per propagation path is governed by the ray-tracing reflection depth, enabling controlled variation of multipath richness across simulation runs. C. The S-ICDF Dataset This section presents the S-ICDF dataset, which is designed to facilitate a detailed assessment of how individual signal and channel parameters influence IQ signal structure and the performance of signal localization algorithms. a) Transmitter & Receiver Setup: A single signal source is placed at a fixed location within the scene and transmits at a center frequency of 1.57542 GHz, corresponding to the GPS- L1 band center frequency. The transmitted waveform is deter- mined by a configurable source-signal parameter and can be selected from multiple signal types. The receiver is modeled as a uniform antenna array with either a 2×2 planar geometry or a 8×1 linear geometry. During acquisition, the array traverses the hall while maintaining a fixed orientation relative to the transmitter. Inter-element spacing is specified by the antenna- distance parameter, and the element gain characteristics are defined by the selected antenna gain pattern. The received signal at each antenna element is sampled over a simulation bandwidth of 100 MHz, resulting in 1,024 complex IQ samples per snapshot. Fig. 3: Trajectory of the antenna. b) ParameterSimulation Methodology:Thedatasetis generated according to a one-at- a-time (OAT) parameter-variation strategy. In each experiment, only a single parameter of interest is varied over its predefined range, while all remaining parameters are fixed at their respective default values. This design enables a clear andunambiguousattributionof performance changes to the parameter under investigation, without confounding effects arising from simultaneous multi-parameter variation. An overview of all simulation, TABLE I: Overview of all simulation, interference, and antenna receiver parameters. Simulation Environment Parameters ParameterValue Bandwidth of simulated IQ samples100 MHz Center frequency of simulated IQ samples1.57542 GHz Received signal length1,024 Number of simultaneous interference sources1 Trajectory length of each recording (number of samples)216,089 Interference & Antenna Setting Parameters ParameterRangeDefault Source bandwidth1 MHz to 20 MHz20 MHz SNR−20 dB to 20 dB20 dB Source signalChirp, Modulated, Noise,Chirp Multitone, FrequHopper Antenna distance[0.05 m, 0.09 m, 0.095 m, 0.2 m]0.09 m Reflection depth[2, 5, 7]5 Refraction[True, False] True Array layout[2× 2 or 8× 1]2× 2 Antenna gain pattern[dipole, iso, hw-dipole, tr38901]dipole interference, and antenna parameters is provided in Table I. For each parameter variation, we generate a complete positioning dataset comprising 216,089 samples, which is subsequently divided into an 80/20 train–test split. Each sample corresponds to a specific point along a receiver trajectory, as illustrated in Figure 3. c) Parameter Description: The following parameters are varied individually over their respective ranges, while all remaining parameters are fixed at their default values, as summarized in Table I. The antenna element spacing is evaluated at four discrete values, namely 0.05 m, 0.09 m, 0.095 m, and 0.2 m, with a default value of 0.09 m. At a center frequency of 1.57542 GHz, this corresponds to inter- element spacings ranging from 0.26λ to 1.1λ. We chose both 0.09 m and 0.095 m as we have observed ambiguities for a wavelength of exactly 0.5λ. For the array layout, two configurations are considered: a 2×2 rectangular array and an 8×1 linear array. These configurations provide two fundamen- tally different aperture geometries, thereby affecting angular resolution and spatial diversity. Four antenna gain patterns are considered. The isotropic pattern serves as an idealized baseline with uniform sensitivity in all directions. The dipole and half-wave dipole patterns introduce directional sensitivity characteristic of practical antenna elements. In addition, the 3GPP TR 38.901 model is included as a standardized antenna pattern. The signal-to-noise (SNR) parameter controls the level of additive noise applied to the received signal and is varied from −20 dB to 20 dB in steps of 2 dB. Low SNR values correspond to severely degraded reception conditions, whereas high SNR values represent near-ideal conditions. Figure 4a shows the histogram of receiver signal power (mean root 2.55.07.510.012.515.0 Signal Strength [MRS] 200 400 600 800 1000 1200 1400 Counts [1000x] Train Test (a) Signal strength. Noise Chirp Frequency- Hopper Modulated Multitone Pulsed 0 2 4 Counts 1e6 Train Test (b) Interference types. 010203040 Bandwidth [MHz] 0 1 2 3 4 Counts 1e6 Train Test (c) Bandwidths. Fig. 4: Histograms of signal sharacteristics. 0246810 Time [μs] -50 -25 0 25 50 Frequency [MHz] (a) Noise. 0246810 Time [μs] -50 -25 0 25 50 Frequency [MHz] (b) Chirp. 0246810 Time [μs] -50 -25 0 25 50 Frequency [MHz] (c) Hopper. 0246810 Time [μs] -50 -25 0 25 50 Frequency [MHz] (d) Modul. 0246810 Time [μs] -50 -25 0 25 50 Frequency [MHz] (e) Multit. 0246810 Time [μs] -50 -25 0 25 50 Frequency [MHz] (f) Pulsed. Fig. 5: Spectrograms for six main interference modulations. squared over IQ samples). The reflection depth determines the maximum number of reflections that a propagation path may undergo within the Sionna ray-tracing engine and is evaluated for values of 2, 5, and 7. Larger values result in a denser and more complex multipath environment, thereby increasing simulation realism at the expense of computational cost. Refraction is modeled as a Boolean parameter that enables or disables refractive propagation effects. When set to True, signals are permitted to propagate through penetrable surfaces, yielding additional propagation paths; when set to False, only reflective propagation paths are considered. This parameter isolates the impact of refractive multipath compo- nents on localization performance. d) Signal Modulation: The source signal defines the transmitted waveform; here, we use predefined generated vector signals rather than simple synthetic waveforms. In total, 102 signal files are included, covering six distinct signal classes and thus a broad range of interference waveform types. Figure 5 shows examples of the six main interference modula- tions, while Figure 4b presents their distribution in the dataset. For Chirp signals, we consider linear (Lin) and parabolic (Para) sweep profiles with fast (Fa) (0.025 ms), medium (Med) (0.25 ms), and slow (Sl) (2.5 ms) sweep times, and vary the bandwidth from 2 to 20 MHz. FrequencyHopper rapidly switch across discrete carrier frequencies, with bandwidth and dwell time as the main varying parameters. Noise signals are spectrally flat emissions with varying occupied bandwidth; ad- ditionally, binary offset carrier (BOC)-modulated variants are included to impose characteristic spectral structure. For Multi- tone, multiple equally spaced sinusoidal tones are transmitted simultaneously, with the number of tones and total occupied bandwidth determining the spectral density. For Modulated, modulation scheme, symbol rate, and occupied bandwidth are varied to capture diverse structured spectral/temporal charac- teristics. Finally, Pulsed signals alternate between transmission and silence with a fixed 1:1 signal-to-silence ratio, while pulse repetition interval and bandwidth are varied. A histogram of the interference bandwidths is given in Figure 4c. 012345 Classes [Prediction] 0 1 2 3 4 5 Classes [Ground Truth] 100.0%0.0%0.0%0.0%0.0%0.0% 0.0%99.7%0.3%0.0%0.0%0.0% 0.0%0.0%100.0%0.0%0.0%0.0% 0.0%0.1%0.1%99.8%0.0%0.0% 0.0%0.0%0.0%0.0%100.0%0.0% 0.0%0.0%0.0%0.0%0.0%100.0% 0 20 40 60 80 100 Percentage [%] 020406080100 Classes [Prediction] 0 20 40 60 80 100 Classes [Ground Truth] Chirp [2]Frequency-Hopper [3]Modulated [4]Multitone [5]Noise [1]Pulsed [6] Chirp [2] Frequency-Hopper [3] Modulated [4] Multitone [5] Noise [1] Pulsed [6] 0 20 40 60 80 100 Percentage [%] Fig. 6: Evaluation of classification (left) and characterization (right) tasks with XceptionTime [27] on raw IQ samples. IV. EXPERIMENTS a) DF with Classical Methods: The primary objective of this dataset is to assess the impact of signal parameters on DF performance. To this end, we establish a baseline using well- known classical methods, namely MUSIC [28], ESPRIT [29], and CAPON [30]. The DF algorithms are evaluated on the 20% test portion of the simulated data, and the mean azimuth and elevation errors are used as benchmark metrics. For MUSIC and ESPRIT, the evaluation grid spans 0 ◦ to 360 ◦ in azimuth and 0 ◦ to 90 ◦ in elevation, with an angular resolution of 1 ◦ in both dimensions. b) Characterization & DF with ML: Our objective is to characterize all 102 interference modulation types. For each of the 4× 6 circle-grid trajectories, we simulate distinct interference parameterizations. We partition the dataset into an 80/20 train–test split and train an XceptionTime [27] model on raw IQ inputs of size 1,024× 4× 2, where 1,024 is the time dimension, 4 is the array dimension, and 2 is the real and imaginary part of the IQ-signal. The dataset comprises 440,821 training samples and 110,172 test samples. The model is trained to predict the interference characteristics, namely the main interference class, modulation type, and bandwidth, using a cross-entropy loss, and to regress the azimuth and elevation angles as well as the source–receiver distance using a mean squared error (MSE) loss. The network outputs a 128- dimensional feature representation that is fed into a linear prediction head. We use a batch size of 64, train for 100 epochs, and use the standard SGD optimizer with an initial learning rate of 10 −4 , applying a multi-step schedule that reduces the learning rate by a factor of 0.1 after 60 and 80 epochs. The overall objective is L total = λ 1 L class + λ 2 L char + λ 3 L az + λ 4 L el + λ 5 L dis , (1) where we set λ 1 = λ 2 = 1, and λ 3 = λ 4 = λ 5 = 0.3. V. EVALUATION Consistently, we report the MSE of relative position (in m), azimuth (in ◦ ), and elevation (in ◦ ), and the accuracy (in %) of classification and characterization. a) Modulation Characterization: Figure 6 shows the results for interference classification and characterization. At the class level, which leads to an accuracy of 99.89%, Noise, Lin,Fa,BW02Lin,Fa,BW05Lin,Fa,BW10Lin,Fa,BW15Lin,Fa,BW20 Lin,Med,BW02Lin,Med,BW05Lin,Med,BW10Lin,Med,BW15Lin,Med,BW20 Lin,Sl,BW02Lin,Sl,BW05Lin,Sl,BW10Lin,Sl,BW15Lin,Sl,BW20 Para,Fa,BW02Para,Fa,BW05Para,Fa,BW10Para,Fa,BW15Para,Fa,BW20 Para,Med,BW02Para,Med,BW05Para,Med,BW10Para,Med,BW15Para,Med,BW20 Para,Sl,BW02Para,Sl,BW05Para,Sl,BW10Para,Sl,BW15Para,Sl,BW20 0.0 2.5 5.0 7.5 10.0 Azimuth [°] ESPRIT MUSIC CAPON (a) Chirp. BW5,2.5B10B15 B10B15 BW0.1BW0.5 BW01BW02BW04 BW05,B 10BW05,B 15 BW05 BW10,AB 15 BW10,B 10 BW10 BW2.5,B 15 BW20,AB 15 BW20 0 2 4 6 8 10 Angle Error [°] ESPRIT MUSIC CAPON (b) Noise. BW01,T10 BW01,T1 BW04,T10 BW04,T1 BW05,B15 ,T10 BW05,B15 ,T1 BW0.2,T10 BW0.2,T1 BW10,AB 15 ,T10 BW10,AB 15 ,T1 BW10,T10 BW10,T1 BW2.5,B15 ,T10 BW2.5,B15 ,T1 0 2 4 6 8 10 Angle Error [°] ESPRIT MUSIC CAPON (c) Pulsed. ABOC 15 BOCc 10BOCc 15 BOCs 10 BOCs 1 BPSK10 BPSK1BPSK5 0 2 4 6 8 10 Angle Error [°] ESPRIT MUSIC CAPON (d) Modulated. Fig. 7: Results of the azimuth (dots) and elevation (cross) orientation errors (in ◦ , MSE) for different interference modulations. “BW” denotes the bandwidth, “B” denotes the width of the BOC separation, “T” denotes the pulse rate in ms. SNR: -20SNR: -18SNR: -16SNR: -14SNR: -12SNR: -10 SNR: -8SNR: -6SNR: -4SNR: -2 SNR: 0SNR: 2SNR: 4SNR: 6SNR: 8 SNR: 10SNR: 12SNR: 14SNR: 16SNR: 18SNR: 20 10 0 10 1 Angular Error [°] ESPRIT MUSIC CAPON Fig. 8: Evaluation results of the azimuth orientation error (in ◦ , MSE, logarithmic) for various SNR [in dB]. Multitone, and Pulsed are identified with high accuracy, as reflected by the strong diagonal structure. Most remaining errors occur between Chirp and FrequencyHopper, whose signatures become similar at small bandwidths. The Modulated class is also occasionally confused with these two categories, although only to a limited extent. For the characterization task, which leads to an accuracy of 71.99%, the majority of errors occur within the same interference types rather than across different classes. In particular, Chirp signals are mainly confused between linear and parabolic sweeps and between medium and slow sweep rates, while FrequencyHopper signals are primarily misclassified across similar bandwidth settings. Within the Noise category, AltBOC and BOC variants exhibit similar spectral patterns, which leads to additional ambiguity. b) Classical DF: Figure 7 compares the DF performance of MUSIC, ESPRIT, and CAPON across interference mod- ulations. Overall, ESPRIT achieves the most accurate and stable azimuth estimates, consistently outperforming MUSIC and CAPON across most waveform variants. The clearest advantage is observed for Chirp and Modulated signals, where ESPRIT maintains comparatively low azimuth errors while MUSIC and CAPON exhibit larger fluctuations. For Noise signals, the performance gap between the methods is smaller, indicating that these waveforms are generally easier to local- ize. In contrast, Pulsed signals appear to be the most chal- lenging, as all methods show increased variability and higher errors for several parameterizations. A further observation is that azimuth errors are generally lower than elevation errors, suggesting that azimuth estimation is less strongly affected by waveform-dependent ambiguities and multipath. In summary, the figure indicates that the interference modulation has a noticeable impact on DF accuracy, with ESPRIT providing the most robust overall performance. TABLE I: Results for different Sionna parameter settings. We evaluate azimuth (ε a ) and elevation (ε e ) errors (in ◦ , MSE). MUSICESPRITCAPON Parameterε a ε e ε a ε e ε a ε e Default parameter setting: Array layout: 2×2, Gain: dipole, Reflection depth = 5, Refraction = False, Spacing: 0.09 m 0.84 2.26 0.78 2.27 0.82 2.23 Array layout: 8×113.19- 13.19- 2.60- Gain: iso0.83 2.26 0.76 2.26 0.81 2.22 Gain: hwdipole0.85 2.27 0.79 2.27 0.83 2.23 Gain: TR 38.9011.86 3.51 1.46 3.51 1.62 3.46 Reflection depth = 22.35 3.53 1.70 3.61 2.08 3.43 Refraction = True0.75 1.98 0.67 1.98 0.74 1.96 Antenna distance: 0.05 m1.22 47.95 1.16 5.41 1.20 5.33 Antenna distance: 0.095 m 1.66 7, 45 0.78 2.27 1.57 2.23 Antenna distance: 0.2 m 102.31 32.56 77.13 61.45 71.48 29.96 c) Evaluation of SNR: Figure 8 shows the azimuth DF error of MUSIC, ESPRIT, and CAPON as a function of the SNR over the range [−20 dB, 20 dB]. The SNR considered here varies within a single environment, as path-loss effects are explicitly accounted for in the simulation. A clear threshold behavior can be observed: below −8 dB, the azimuth error increases strongly for all methods, whereas above −8 dB the error decreases rapidly and saturates at a low value. For low- SNR, additive noise increasingly masks the phase and ampli- tude relations between antenna elements, which degrades the covariance-matrix estimate and makes subspace separation less reliable. As the SNR increases, the signal component becomes more dominant, leading to more stable direction estimates and lower azimuth errors. At high SNR, the remaining error is mainly limited by factors such as multipath propagation, array geometry, and finite sample effects rather than by noise. d) Evaluation of Sionna Parameters: Table I shows that the dataset captures meaningful dependencies on key parameters. Under default settings, all three methods perform well, with azimuth errors around 0.8 ◦ and elevation errors around 2.2 ◦ . ESPRIT consistently outperforms MUSIC and CAPON. The antenna gain pattern has a moderate effect: dipole, isotropic, and half-wave dipole yield nearly identical results, whereas the 3GPP TR 38.901 pattern noticeably de- grades both azimuth and elevation accuracy. A higher reflec- tion depth leads to a more robust DF, while enabling refraction slightly improves all methods, likely because propagation paths provide more stable spatial information. The default antenna spacing of 0.09 m gives the best overall performance. At 0.095 m MUSIC and CAPON have spatial ambiguities 1015202530 x [m] 8 10 12 14 16 18 20 22 24 y [m] Azimuth: 1.32° Jammer 0 5.0 60.0 Azimuth Error [°] (a) Chirp. 1015202530 x [m] 8 10 12 14 16 18 20 22 24 y [m] Azimuth: 1.44° Jammer 0 5.0 60.0 Azimuth Error [°] (b) FrequHopper. 1015202530 x [m] 8 10 12 14 16 18 20 22 24 y [m] Azimuth: 1.96° Jammer 0 5.0 60.0 Azimuth Error [°] (c) Noise/Pulsed. Fig. 9: Evaluation of the azimuth prediction error. between the estimated directions. Similarly, for the 8 × 1 array configuration with a normalized element spacing of 0.5λ, spatial ambiguities are observed, where the estimated source direction is occasionally mirrored at broadside. To account for this symmetric ambiguity, azimuth estimates for the 8×1 array are normalized to [−90 ◦ , 90 ◦ ]. e) ML-Based DF: The final DF performance of the XceptionTime model achieves a mean error of 1.85 ◦ in az- imuth, 2.33 ◦ in elevation, and 0.26 m in position. Figure 9 illustrates the azimuth prediction error across all trajectory points for three representative interference types. Overall, the azimuth error tends to increase when the receiver is located closer to the interference source. Among the shown examples, Noise and Pulsed interference yields the highest azimuth error at 1.96 ◦ , whereas Chirp interference results in the lowest error at 1.32 ◦ , indicating greater robustness of the model for this waveform class. This trend is consistent with the observations made for the classical DF methods. VI. CONCLUSION We introduced S-ICDF, a large-scale Sionna-based dataset for interference characterization and DF with controlled pa- rameter variation. The benchmark demonstrates that S-ICDF captures meaningful dependencies on waveform, SNR, mul- tipath, and antenna design, enabling systematic sensitivity analysis of both classical and ML-based methods. In our evaluation, XceptionTime achieves 99.89% interference clas- sification accuracy and competitive DF performance, while ESPRIT provides the most robust classical baseline (0.78 ◦ ). REFERENCES [1] R. M. Ferre, A. de la Fuente, and E. S. Lohan, “Jammer Classification in GNSS Bands via Machine Learning Algorithms,” in MDPI Sensors, vol. 19(22), Nov. 2019. [2] C. J. Swinney and J. C. Woods, “GNSS Jamming Classification via CNN, Transfer Learning & the Novel Concatenation of Signal Repre- sentations,” in CyberSA, Dublin, Ireland, Jun. 2021. [3] H. Pirayesha and H. Zeng, “Jamming Attacks and Anti-Jamming Strate- gies in Wireless Networks: A Comprehensive Survey,” in COMST, vol. 24(2), Mar. 2022, p. 767–809. [4] I. E. Mehr, A. Minetto, and F. Dovis, “A Navigation Signals Monitoring, Analysis and Recording Tool: Application to Real-Time Interference Detection and Classification,” in ION GNSS+, Sep. 2023. [5] N. R. S. Miguel, Y.-H. Chen, S. Lo, T. Walter, and D. Akos, “Calibration of RFI Detection Levels in a Low-Cost GNSS Monitor,” in IEEE/ION PLANS, Monterey, CA, Apr. 2023. Acknowledgments. This work has been carried out within the PaiL project, funding code 50NP2506, sponsored by the German Federal Ministry for Transport (BMV) and supported by the German Space Agency at DLR, the Bundesnetzagentur (BNetzA), and the Federal Agency for Cartography and Geodesy (BKG). [6] F. Ott, L. Heublein, N. L. Raichur, T. Feigl, J. Hansen, A. R ̈ ugamer, and C. Mutschler, “Few-Shot Learning with Uncertainty-based Quadruplet Selection for Interference Classification in GNSS Data,” in ICL-GNSS, Antwerp, Belgium, Jun. 2024. [7] L. Heublein, T. Feigl, T. Nowak, A. R ̈ ugamer, C. Mutschler, and F. Ott, “Evaluating ML Robustness in GNSS Interference Classification, Characterization & Localization,” in ICL-GNSS, Rome, Italy, Jun. 2025. [8] N. L. Raichur, T. Brieger, D. Jdidi, T. Feigl, J. R. van der Merwe, B. Ghimire, F. Ott, A. R ̈ ugamer, and W. Felber, “Machine Learning- assisted GNSS Interference Monitoring Through Crowdsourcing,” in ION GNSS+, Denver, CO, Sep. 2022, p. 1151–1175. [9] L. Heublein, C. Wielenberg, T. Nowak, T. Feigl, C. Mutschler, and F. Ott, “Attention-Based Fusion of IQ and FFT Spectrograms with AoA Features for GNSS Jammer Localization,” in RadarConf, Krakow, Poland, Oct. 2025. [10] F. Chen, Z. Liu, L. Huang, Y. Xie, B. Ren, and Q. Zhou, “GNSS Interference Mitigation Method Based on Deep Learning,” in Sec. Interdisciplinary Physics, Mar. 2025. [11] J. H. Yang, C. H. Kang, S. Y. Kim, and C. G. Park, “Intentional GNSS Interference Detection and Characterization Algorithm Using AGC and Adaptive IIR Notch Filter,” in IJASS, vol. 13(4), 2012, p. 491–498. [12] M. J. Murrian, L. Narula, and T. E. Humphreys, “Characterizing Terrestrial GNSS Interference from Low Earth Orbit,” in ION GNSS+, Miami, Florida, Sep. 2019, p. 3239–3253. [13] J. N. Gross and T. E. Humphreys, “GNSS Spoofing, Jamming, and Mul- tipath Interference Classification Using a Maximum-Likelihood Multi- tap Multipath Estimator,” in ION GNSS+, Jan. 2017. [14] N. L. Raichur, L. Heublein, T. Feigl, A. R ̈ ugamer, C. Mutschler, and F. Ott, “Bayesian Learning-driven Prototypical Contrastive Loss for Class-Incremental Learning,” in TMLR, Apr. 2025. [15] L. Heublein, T. Feigl, A. R ̈ ugamer, C. Mutschler, and F. Ott, “Variational and Generative Models with Quantization for Disentanglement and Compressed Sensing of GNSS Spectrograms,” in J-ISPIN, Jan. 2026. [16] H. He and E. A. Garcia, “Learning from Imbalanced Data,” in TKDE, vol. 21(9), Jun. 2009, p. 1263–1284. [17] F. C. Commission, “Jammer Enforcement,” Apr. 2020. [Online]. Available: https://w.fcc.gov/general/jammer-enforcement [18] A. Broumandan, A. Pirsiavash, I. Tremblay, and S. Kennedy, “Eval- uation of NovAtel’s Jamming and Spoofing Detection and Mitigation Capabilities During Jammertest2024,” in ION GNSS+, Jan. 2025. [19] T. Humphreys, J. Bhatti, D. Shepard, and K. Wesson, “The Texas Spoofing Test Battery: Toward a Standard for Evaluating GPS Signal Authentication Techniques,” in ION GNSS+, Sep. 2012. [20] C. J. Swinney and J. C. Woods, “Raw IQ Dataset for GNSS GPS Jamming Signal Classification,” in Zenodo, Mar. 2021. [21] S. Rahman, M. Z. H. Bhuiyan, J. Nurmi, and E. Lohan, “Descriptor: Tuni2025 GNSS - Galileo and GPS Spoofing Datasets (TG-GGSD),” in IEEE Data Descriptions, Oct. 2025. [22] R. M. Ferre, P. Richter, A. D. L. Fuente, and E. S. Lohan, “In-lab Validation of Jammer Detection and Direction Finding Algorithms for GNSS,” in ICL-GNSS, Nuremberg, Germany, Jun. 2019. [23] P. Richter, R. M. Ferre, and E.-S. Lohan, “GATEMAN Project – Wide- Bandwidth, High-Precision GNSS and Jammer Raw Data,” in Zenodo, Apr. 2019. [24] D. G ́ omez-Casco, P. Crosta, and M. Spangenberg, “Validation of Evil WaveForms in a GNSS Simulator for GPS and Galileo Signals,” in NAVITEC, Noordwijk, Netherlands, Apr. 2022. [25] A. Albright, S. Powers, J. Bonior, and F. Combs, “Oak Ridge Spoofing and Interference Test Battery (OAKBAT) - GPS,” in U.S. Department of Energy, Sep. 2020. [26] J. Hoydis, S. Cammerer, F. A. Aoudia, A. Vem, N. Binder, G. Marcus, and A. Keller, “Sionna: An Open-Source Library for Next-Generation Physical Layer Research,” in arXiv:2203.11854, Mar. 2022. [27] E. Rahimian, S. Zabihi, S. F. Atashzar, A. Asif, and A. Mohammadi, “XceptionTime: A Novel Deep Architecture based on Depthwise Separa- ble Convolutions for Hand Gesture Classification,” in arXiv:1911.03803, Nov. 2019. [28] R. Schmidt, “Multiple Emitter Location and Signal Parameter Estima- tion,” in TAP, vol. 43(3), Mar. 1986, p. 276–280. [29] R. Roy and T. Kailath, “ESPRIT – Estimation of Signal Parameters via Rotational Invariance Techniques,” in TAES, Jul. 1989. [30] A. Hassanien, S. Shahbazpanahi, and A. B. Gershman, “A Generalized Capon Estimator for Localization of Multiple Spread Sources,” in IEEE TSP, vol. 52(1), Jan. 2024.