As traditional Euclidean, Transformer-based classifiers (e.g., ET-BERT) become increasingly vulnerable to adversarial pre-padding and cryptographic mimicry (e.g., VLESS Reality), the future of encrypted traffic classification relies on non-Euclidean flow physics.

This dataset represents a massive 40:1 compression of over 400GB of raw, real-world network traffic into 11GB of highly optimized PyTorch tensors. By discarding payload bytes entirely, this dataset focuses purely on continuous-time thermodynamic topology, enabling the training of Liquid State Space Models (SSMs), Mamba architectures, and ODE-based neural networks for zero-day evasion detection.

🧬 The 6-Dimensional Physics Vector

Each tensor in this dataset represents a causal sequence of 1,000 packets (Shape: [1000, 6]). The 6 extracted features are entirely payload-independent:

Packet Volume (Size): The absolute byte size of the packet.
Inter-Arrival Time (IAT): The microsecond-resolution temporal gap between packets, used to track the thermodynamic decay of automated routing.
Directionality: Egress (+1) or Ingress (-1) to establish causal flow asymmetry.
TCP Window Size: The advertised receiver window, capturing algorithmic congestion control artifacts.
TCP Flags: Stateful protocol metadata normalized as a continuous scalar.
Payload Ratio: The ratio of payload bytes to total frame size, tracking the morphological overhead of adversarial padding.

⚔️ The 4-Tier Threat Architecture

The dataset is structured under a strict Zero-Trust paradigm, aggregating traffic into four distinct tiers:

Tier I: Planetary Baselines (Benign - Class 0)
- Sources: MAWI WIDE Project (2026), CIC-IDS-2017, Corporate DNS captures.
- Profile: High stochastic variance, natural human network jitter, standard TCP congestion curves.
Tier II: Automated Swarm (Malicious - Class 1)
- Sources: Aposemat IoT-23 (Mirai, Torii), CTU-13.
- Profile: Hardware captures of IoT botnets exhibiting highly rigid IAT and automated sequence timing.
Tier III: Zero-Day Vault (Malicious - Class 1)
- Sources: Malware-Traffic-Analysis (2026), BCCC-Mal-NetMem.
- Profile: Memory-resident rootkits, Emotet, ClickFix, and high-variance C2 beaconing.
Tier IV: Cryptographic Evasion (Malicious - Class 1)
- Sources: Vixero Custom Red Team (VLESS Reality, GhostBear, AMOI Morphing).
- Profile: Proprietary evasion tunnels designed to shatter Euclidean manifolds and anchor volumetric data to benign distributions.

🔒 Privacy & Ethical Disclosure

Zero PII. Zero Payload Bytes. To ensure strict ethical compliance and privacy preservation, all raw payloads, IP addresses, MAC addresses, and cryptographic certificates have been stripped prior to tensor generation. This dataset contains purely mathematical flow physics.

Note: The raw PCAPs for Tier IV remain gated and closed-source to prevent the proliferation of proprietary evasion protocols to active threat actors.

💻 Usage (PyTorch)

Loading a sequence into your model is native and hyper-fast:

import torch

# Load a single 1000-packet causal sequence
sequence_tensor = torch.load('data/tier2_mirai_seq_001.pt')

# Verify shape: [Sequence_Length, Features] -> [1000, 6]
print(sequence_tensor.shape)

✒️ Citation & Author

Created and maintained by Vickson Ferrel (Vixero Technology Enterprise). If you utilize this dataset in your research for continuous-time sequence modeling or thermodynamic variance detection, please refer to the AEGIS pre-print on arXiv (cs.CR / cs.LG).

ORCID: 0009-0009-0155-0913 | Web: vixdev.cloud

Downloads last month: 13

Total file size:

11 GB