Spectrum Anomaly Detection in OFDMA Systems: Simulation Framework and Benchmark Dataset

Overview This dataset accompanies the paper titled Spectrum Anomaly Detection in OFDMA Systems: Simulation Framework and Benchmark Dataset and serves as a benchmark for spectrum anomaly detection methods. It is generated through simulations of an OFDMA wireless network in an industrial indoor environment, with signal propagation modeled via ray tracing. The dataset consists of 20,000 samples, where each sample includes spectrograms from 21 spatially distributed sensing units (SUs), binary resource allocation maps, and metadata such as jammer type, signal-to-noise ratio (SNR), and signal-to-jammer ratio (SJR). Half of the samples represent normal conditions, while the other half contain anomalies caused by five distinct types of jammers, including simple noise-like and sophisticated deceptive jamming attacks. The dataset is intended for both supervised and unsupervised machine learning approaches and provides baseline implementations and evaluations. This comprehensive dataset enables standardized, reproducible evaluations of spectrum anomaly detection algorithms and supports further research into resilient wireless communication systems. Related links Preprint: https://arxiv.org/abs/2606.02102 GitHub: https://github.com/akdd11/ofdma-spectrum-anomalies-simulation Dataset Structure The dataset comprises overall 20000 samples, whereby each sample denotes one realization of transmitter and, if applicable, jammer locations and signals. Half of those samples belong to the normal class, and the other half to the type anomaly. The anomaly samples are distributed evenly over five jammer types (see the paper for details), i.e., there are 2000 samples for each jammer type. Spectrograms Since a set of 21 spatially distributed sensing units (SUs) acting as receivers is considered, each sample consists of 21 spectrograms that are provided as 8-bit grayscale PNG images. Min-max scaling is applied over all images in the dataset. To apply inverse scaling, the minimum and maximum values of the overall dataset are provided in spectrogram_min_max.csv. Files are named spectrogram-<sample_index>-<su_index>.png. Resource allocation information Moreover, for each sample, the resource blocks that are allocated to legitimate users are provided as 8-bit grayscale PNG images. Those images might be used as additional input features to further enhance the detection capability. Files are named alloc_res-<sample_index>.png. Note: The pixel value corresponds to the transmitter index (1-10), with 0 indicating "not allocated". Due to the small values (compared to the maximum pixel value of 255), the images seem almost completely black. Yet, the information is contained. The repository linked above contains a notebook to load the images and visualize them in a discrete color map, where the allocation can be easily recognized. Yet, for ML training, the grayscale images seem more appropriate than discrete RGC color maps. Labels The labels.csv contains metadata for each sample, comprising the following features: # Column name Feature 1 jammer_type Jammer type [string] (empty for normal samples) 2 jammer_power Transmit power of the jammer in dBm [float] (empty for normal samples) 3 jammer_location Jammer location in m [array] (empty for normal samples) 4 num_legitimate_transmitters Number of legitimate transmitters [int] 5 - 25 snr_by_su_<su_idx> SNR per SU (spectrogram) in dB [float] 26 - 46 sjr_by_su_<su_idx> SJR per SU (spectrogram) in dB [float] (inf for normal samples)