SliceVision-F2I: A Synthetic Feature-to-Image Dataset for Visual Pattern Representation on Network Slices

Md. Abid Hasan Rafi, Mst. Fatematuj Johora, Pankaj Bhowmik · 2025

Read SliceVision-F2I: A Synthetic Feature-to-Image Dataset for Visual Pattern Representation on Network Slices by Md. Abid Hasan Rafi, Mst. Fatematuj Johora, Pankaj Bhowmik (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-01087.txt.

Introduction

Md. Abid Hasan Rafi, Mst. Fatematuj Johora, and Pankaj Bhowmik release SliceVision-F2I, a synthetic data set for studying network slices through pictures. Network slicing, central to 5G and 6G services, needs reliable ways to identify slice states, and visual methods need data to train and test on.

Each sample pairs a vector of network performance indicators with a color image encoding it. Four encoding schemes are used, based on physical mappings, Perlin noise, neural wallpapering, and fractal branching, with thirty thousand samples per scheme at low resolution. Noise and realistic imperfections are added deliberately. The data set suits visual learning, network state classification, anomaly detection, and benchmarking of image based methods on network data, and it is publicly available.

Catalog record

CREATORMd. Abid Hasan Rafi, Mst. Fatematuj Johora, Pankaj Bhowmik
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE23 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-01087

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

SliceVision-F2I: A Synthetic Feature-to-Image Dataset for Visual Pattern Representation on Network Slices Md. Abid Hasan Rafi, Mst. Fatematuj Johora, Pankaj Bhowmik Abstract The emergence of 5G and 6G networks has established network slicing as a significant part of future service-oriented architectures, demanding refined identification methods supported by robust datasets. The article presents SliceVision-F2I, a dataset of synthetic samples for studying feature visualization in network slicing for next-generation networking systems. The dataset transforms multivariate Key Performance Indicator (KPI) vectors into visual representations through four distinct encoding methods: physically inspired mappings, Perlin noise, neural wallpapering, and fractal branching. For each encoding method, 30,000 samples are generated, each comprising a raw KPI vector and a corresponding RGB image at low-resolution pixels. The dataset simulates realistic and noisy network conditions to reflect operational uncertainties and measurement imperfections. SliceVision-F2I is suitable for tasks involving visual learning, network state classification, anomaly detection, and benchmarking of image-based machine learning techniques applied to network data. The dataset is publicly available and can be reused in various research contexts, including multivariate time series analysis, synthetic data generation, and feature-to-image transformations. K eywords Network Slicing \cdot Procedural Generation \cdot Machine Learning \cdot Visual encoding \cdot Network KPI Simulation 1 Introduction The advent of 5G and 6G networks has elevated network slicing from a conceptual framework to a fundamental architectural paradigm [ 1 ] . As noted by [ 2 ] , this transformation demands intelligent management systems capable of handling the inherent complexity of multi-service environments. However, current approaches face significant challenges in achieving reliable slice classification under real-world conditions characterized by measurement noise and operational uncertainties [ 3 ] . Traditional machine learning methods operating on raw Key Performance Indicators (KPIs) have demonstrated limitations in this context. While hybrid deep learning approaches like those proposed by [ 4 ] achieve impressive accuracy in controlled settings, their performance degrades when confronted with the imperfect measurements typical of operational networks. Recent work has shown that visual representations of network data can overcome some of these limitations [ 5 ] , building on earlier insights about the value of visual analytics in complex systems [ 6 ] . However, the lack of standardized datasets for visual network analytics has hindered progress in this promising direction [ 7 ] . This work presents SliceVision-F2I [ 8 ] , an innovative dataset that tackles these difficulties with four principal contributions: • A synthetic dataset of 30,000 samples per generation method combining raw KPI vectors with corresponding visual encodings, specifically designed to reflect realistic noisy conditions. • Four distinct visual transformation methods (physically inspired mappings, Perlin noise, neural wallpapering, and fractal branching) enable perfect classification accuracy via neural networks. • Demonstration that low-resolution visual representations maintain classification performance enabling potential real-time processing. • A benchmark for comparing traditional ML approaches with vision-based methods in network slice identification. Although our current results achieve perfect classification accuracy, the dataset opens numerous avenues of research beyond this achievement. First, the visual encoding methodology itself represents a novel contribution that others can extend to different network analytics tasks. Second, [ 5 ] explain that these types of visual model may make AI systems for network management more straightforward to understand. Lastly, the dataset gives us a consistent place to try new classification methods that are strong enough to deal with the noise and change that come with real-world networks. A brief overview of the dataset is provided in Table 1 . Table 1: SliceVision-F2I Dataset Specifications Category Description Domain AI/Networking for 5G/6G Network Slicing and Visual Insights Content 10 KPIs (delay, jitter, loss, throughput, retransmissions, discard, RSSI, SNR, CPU/mem) + 16×16 RGB patterns: physical, Perlin, wallpaper, fractal Generation 30k samples/method (eMBB/URLLC/mIoT); Noise: Gaussian + 5% MAR; Encodings: physical (geometric), Perlin (parameterized), wallpaper (periodic), fractal (recursive) Access Mendeley (DOI: 10.17632/68xp3vszsz.1); Formats: CSV (KPIs), NPY (images) Application Network Management, ML Research, Telecom Analysis The following sections of this work are organized as follows: Section 2 examines pertinent literature on network slicing and visual analytics. Section 3 delineates our process for dataset production. Section 4 outlines the experimental findings, and Section 5 provides concluding observations. 2 Contextual Study The evolution toward service-oriented 5G/6G networks has made network slicing indispensable, enabling diverse Quality of service (QoS) requirements through shared infrastructure [ 1 ] . While AI/ML approaches like hybrid CNN-LSTM models achieve 95.17% accuracy in controlled environments [ 4 ] , their performance degrades under real-world noise and dynamic conditions [ 3 ] . Traditional methods relying on raw KPIs [ 2 ] or transport-layer simulations [ 7 ] lack mechanisms to handle operational uncertainties, creating a critical need for robust feature representations. Recent work demonstrates that visual abstractions - from graph-based KPI relationships [ 6 ] to synthetic network heatmaps [ 5 ] - can capture latent patterns that conventional ML misses. Our work advances this paradigm through three key innovations: • Unlike single-encoding approaches [ 5 ] , we systematically evaluate four distinct visual transformations: physical mappings , Perlin noise , neural wallpapering , and fractal branching for slice identification. • Where prior datasets focus on ideal conditions [ 7 ] , SliceVision-F2I [ 8 ] incorporates realistic noise models that mimic measurement imperfections. • While existing methods require high-resolution inputs [ 6 ] , we demonstrate perfect classification at low resolutions , enabling real-time deployment. This positions visual KPI encoding as both a theoretical breakthrough and a practical solution for noisy operational environments. 3 Dataset Design Figure 1: Directory Hierarchy of SliceVision-F2I Our dataset generation framework transforms multidimensional network slice KPIs into four distinct visual representations, each designed to capture unique structural or physical characteristics of network behavior. The synthetic dataset consists of 30,000 samples per generation method, provided in both NumPy array and CSV formats. It covers three slice types: eMBB, URLLC, and mIoT, with realistic class distributions and embedded noise patterns. An overview of the dataset structure is shown in Figure 1 . 3.1 KPI Simulation Model The network slice performance data set is generated using a multivariate simulation engine that produces 10 key performance indicators (KPIs) for three types of slice. The complete KPI vector is defined as: \mathcal{K}=\{\delta,\jmath,\lambda,\tau,\rho,\psi,r,s,c,m\} (1) where each component represents a critical network metric as detailed in Table 2 . Table 2: KPI Definitions and Units Symbol Description Unit \delta End-to-end delay ms \jmath Packet delay variation (jitter) ms \lambda Packet loss rate % \tau Throughput Mbps \rho Retransmission probability % \psi Packet discard rate % r Received Signal Strength Indicator (RSSI) dBm s Signal-to-Noise Ratio (SNR) dB c CPU utilization % m Memory utilization % Each KPI k_{i} follows slice-specific statistical distributions with controlled variation parameters: k_{i}\sim\mathcal{N}(\mu_{t},\sigma_{t})\cdot\left[1+\mathcal{U}(-\alpha,\alpha)\right],\quad t\in\{\text{eMBB},\text{URLLC},\text{mIoT}\} (2) where \mu_{t} and \sigma_{t} are the slice-type-specific mean and standard deviation values derived from 3GPP TR 28.801, as shown in Table 3 . Table 3: Slice-Type Base Parameters KPI eMBB ( \mu\pm\sigma ) URLLC ( \mu\pm\sigma ) mIoT ( \mu\pm\sigma ) \delta (ms) 10\pm 1.5 0.5\pm 0.075 50\pm 7.5 \jmath (ms) 2\pm 0.3 0.1\pm 0.015 10\pm 1.5 \lambda (%) 1\pm 0.2 0.001\pm 0.0002 5\pm 1.0 \tau (Mbps) 200\pm 20 5\pm 0.5 0.1\pm 0.02 The simulation model incorporates several realistic network conditions through: \mathcal{M}(k_{i})=\begin{cases}\text{NaN}&\text{with probability }\beta\quad\text{(missing data)}\\ k_{i}+\epsilon&\text{otherwise}\end{cases} (3) where \epsilon\sim\mathcal{N}(0,\sigma_{noise}) represents measurement noise

Read the complete text