---
license: cc-by-nc-4.0
library_name: pytorch
tags:
  - cybersecurity
  - network-traffic
  - intrusion-detection
  - tabular-classification
  - synthetic-data
  - xgboost
  - baseline
pipeline_tag: tabular-classification
base_model: []
datasets:
  - xpertsystems/cyb001-sample
metrics:
  - accuracy
  - f1
model-index:
  - name: cyb001-baseline-classifier
    results:
      - task:
          type: tabular-classification
          name: 3-class network flow classification
        dataset:
          type: xpertsystems/cyb001-sample
          name: CYB001 Synthetic Network Traffic (Sample)
        metrics:
          - type: accuracy
            value: 0.9980
            name: Test accuracy (XGBoost)
          - type: f1
            value: 0.9961
            name: Test macro-F1 (XGBoost)
          - type: accuracy
            value: 0.9932
            name: Test accuracy (MLP)
          - type: f1
            value: 0.9869
            name: Test macro-F1 (MLP)
---

# CYB001 Baseline Classifier

**Multi-class network flow classifier trained on the CYB001 synthetic
network traffic sample. Predicts `BENIGN`, `MALICIOUS`, or `AMBIGUOUS`
from per-flow features.**

> **Baseline reference, not for production use.** This model demonstrates
> that the [CYB001 sample dataset](https://huggingface.co/datasets/xpertsystems/cyb001-sample)
> is learnable end-to-end and gives prospective buyers a working starting
> point to evaluate against their own pipelines. It is not an intrusion
> detection system. See [Limitations](#limitations).

## Model overview

| Property | Value |
|---|---|
| Task | 3-class flow classification (BENIGN / MALICIOUS / AMBIGUOUS) |
| Training data | `xpertsystems/cyb001-sample` (9,770 flows, sample only) |
| Models | XGBoost + PyTorch MLP |
| Input features | 101 (after one-hot encoding) |
| License | CC-BY-NC-4.0 (matches dataset) |
| Status | Reference baseline |

Two model artifacts are published. They are designed to be used together — disagreement between them is itself a useful triage signal:

- `model_xgb.json` — gradient-boosted trees, primary recommendation
- `model_mlp.safetensors` — PyTorch MLP in SafeTensors format

## Quick start

```bash
pip install xgboost torch safetensors pandas huggingface_hub
```

```python
from huggingface_hub import hf_hub_download
import json, numpy as np, torch, xgboost as xgb
from safetensors.torch import load_file

REPO = "xpertsystems/cyb001-baseline-classifier"

# Download artifacts
paths = {n: hf_hub_download(REPO, n) for n in [
    "model_xgb.json", "model_mlp.safetensors",
    "feature_engineering.py", "feature_meta.json", "feature_scaler.json",
]}

# Make feature pipeline importable
import sys, os
sys.path.insert(0, os.path.dirname(paths["feature_engineering.py"]))
from feature_engineering import transform_single, load_meta, INT_TO_LABEL

meta = load_meta(paths["feature_meta.json"])
xgb_model = xgb.XGBClassifier(); xgb_model.load_model(paths["model_xgb.json"])

# Predict (see inference_example.ipynb for full single-record example)
X = transform_single(my_flow_record_dict, meta)
proba = xgb_model.predict_proba(X)[0]
print(INT_TO_LABEL[int(np.argmax(proba))])
```

See [`inference_example.ipynb`](./inference_example.ipynb) for a full
copy-paste demo including the MLP load path and a batch run on 200 rows
from the public sample.

## Training data

Trained on the public sample of CYB001, 9,770 flows with:

| Label | Train (n=6,838) | Test (n=1,466) | Test share |
|---|---:|---:|---:|
| BENIGN | 4,916 | 1,054 | 71.9% |
| MALICIOUS | 1,378 | 295 | 20.1% |
| AMBIGUOUS | 544 | 117 | 8.0% |

Split: 70 / 15 / 15 stratified by label, seed 42.

Class imbalance was addressed with `class_weight='balanced'` (XGBoost
`sample_weight`) and weighted cross-entropy (MLP). Stratified splitting
preserves the proportion in each fold.

### Dataset calibration anchors

The CYB001 sample is calibrated to 12 named industry signatures. The
features that surface most prominently in the baseline correspond to
these anchors:

| Calibrated signature | Target | Observed (sample) | Feature(s) the model uses |
|---|---:|---:|---|
| `c2_beacon_regularity_score` | 0.78 | 0.77 | `iat_cv`, `inter_arrival_time_std` |
| `payload_entropy_benign_mean` | 4.80 | 4.86 | `payload_entropy_mean` |
| `fwd_bwd_byte_ratio_benign` | 1.34 | 1.41 | `fwd_bwd_byte_ratio` |
| `malicious_flow_rate` | 0.172 | 0.202 | (class prior) |
| `protocol_violation_rate` | 0.015 | 0.016 | `protocol_violation_flag`, `protocol_violation_count` |
| `scan_probe_density` | 0.043 | 0.045 | `tcp_flag_anomaly_score`, port features |

Full benchmark table in the [dataset card](https://huggingface.co/datasets/xpertsystems/cyb001-sample).

## Feature pipeline

The bundled `feature_engineering.py` is the canonical feature recipe.
The training script and the inference example both call into it.

**Three columns are deliberately excluded** because they leak the label:

- `traffic_category` — perfectly deterministic of label (every `attack_*`
  category is 100% MALICIOUS, etc.).
- `attack_subcategory` — non-null iff label is MALICIOUS.
- `attacker_capability_tier` — generator metadata labeled per flow
  including benign flows; not a real-world observable at inference time.

**Five session-level features were kept** after a per-label leakage audit
(`payload_entropy_mean`, `retransmission_rate`, `protocol_violation_count`,
`c2_beacon_flag`, `session_risk_score`) because their distributions
overlap meaningfully across labels (i.e. they behave like detector
outputs, not oracles). **Three were dropped** (`exfil_volume_bytes`,
`scan_probe_count`, `lateral_move_flag`) because they are zero for all
non-MALICIOUS rows.

Engineered features (each encodes a stated domain hypothesis, see source
for the one-line rationale per feature):

- `iat_cv` — inter-arrival-time coefficient of variation. C2 beacon signature.
- `fwd_bwd_byte_ratio` — exfiltration signature.
- `bytes_per_packet_fwd`, `payload_density` — flow shape.
- `tcp_flag_anomaly_score` — RST/URG/FIN density. Scan and protocol-misuse signature.
- `hour_of_day`, `is_off_hours` — diurnal pattern. APT and insider tiers are off-peak biased in the dataset calibration.
- `is_well_known_dest_port`, `is_ephemeral_src_port` — port observables.

## Evaluation

### Test-set metrics (n = 1,466, stratified)

**XGBoost**

| Metric | Value |
|---|---:|
| Accuracy | 0.9980 |
| Macro-F1 | 0.9961 |
| Weighted-F1 | 0.9980 |
| Macro ROC-AUC (OvR) | ≈ 1.00 |

| Class | F1 | Support |
|---|---:|---:|
| BENIGN | 0.9986 | 1,054 |
| MALICIOUS | 0.9983 | 295 |
| AMBIGUOUS | 0.9915 | 117 |

**MLP**

| Metric | Value |
|---|---:|
| Accuracy | 0.9932 |
| Macro-F1 | 0.9869 |
| Weighted-F1 | 0.9932 |

| Class | F1 | Support |
|---|---:|---:|
| BENIGN | 0.9962 | 1,054 |
| MALICIOUS | 0.9899 | 295 |
| AMBIGUOUS | 0.9746 | 117 |

Confusion matrices and per-class precision/recall are in
[`validation_results.json`](./validation_results.json).

### Ablation: contribution of session-level features

To check whether the model is genuinely reading the flow-level signal or
leaning on session aggregates, the same XGBoost configuration was trained
with all five session-aggregate features removed:

| Configuration | Accuracy | Macro-F1 | AMBIGUOUS F1 |
|---|---:|---:|---:|
| Full feature set (published) | 0.9980 | 0.9961 | 0.991 |
| Flow-only (session aggregates dropped) | 0.9884 | 0.9776 | 0.957 |

The session join contributes about **+1.0 pp** of accuracy and **+0.02**
macro-F1. The model is not session-dominated; the flow-level features
carry the bulk of the signal. The full numbers for both configurations
are in [`ablation_results.json`](./ablation_results.json).

### Architecture

**XGBoost:** multi-class gradient boosting (`multi:softprob`, 3 classes),
`hist` tree method, class-balanced sample weights, early stopping on
validation macro-F1.

**MLP:** `n_features → 128 → 64 → 3`, each hidden layer followed by
`BatchNorm1d` → `ReLU` → `Dropout(0.3)`, weighted cross-entropy loss,
AdamW optimizer, early stopping on validation macro-F1.

Training hyperparameters (learning rate, batch size, n_estimators,
early-stopping patience, weight decay, class-weighting strategy) are
held internally by XpertSystems and are not part of this release.

## Limitations

**This is a baseline reference, not an intrusion detection system.**

1. **Performance is inflated by synthetic structure.** The numbers above
   reflect performance on calibrated synthetic data where the BENIGN and
   attack categories sit on distinct statistical signatures by
   construction. A real production IDS facing live traffic must contend
   with concept drift, adversarial evasion, encrypted-traffic ambiguity,
   and a much fatter long tail of benign behaviour. Expect substantial
   degradation when transferring to real CICIDS-style datasets or
   in-the-wild traffic.

2. **Sample size for `AMBIGUOUS` is small.** Only 117 test examples;
   the per-class F1 has wide confidence bands. The full CYB001 product
   (~62k AMBIGUOUS flows out of ~500k) supports more reliable estimation.

3. **Trained on the public 1/60th sample only.** The full product
   contains additional traffic categories, longer sequences, and
   richer adversary behaviour. A model trained on the full dataset
   would perform differently — likely lower headline accuracy with
   better calibration and generalisation. The intent of this release
   is reference, not state-of-the-art.

4. **Topology features are static labels, not signals.** Fields like
   `defender_architecture` and `firewall_policy` are descriptive
   categorical attributes of the network segment, not learned defender
   responses. They help the model condition on context but do not
   simulate real adversarial dynamics.

5. **MLP brittleness on OOD inputs.** With ~7k training rows, the MLP
   can produce confidently-wrong predictions on hand-crafted records
   whose feature combinations are far from the training manifold. The
   inference notebook demonstrates this. XGBoost is more robust here.
   In practice, use both and treat disagreement as a signal for review.

6. **Class imbalance handling is straightforward.** Class-balanced
   weights work for this sample but production-scale rare-class
   detection (e.g. APT C2 at < 0.1% of traffic) needs more careful
   threshold calibration, ranking metrics, and likely calibrated
   probabilities rather than argmax classification.

## Intended use

- **Evaluating fit** of the CYB001 dataset for your IDS / NDR research
- **Baseline reference** for new model architectures on synthetic
  network traffic
- **Teaching and demo** for tabular classification on flow-level features
- **Feature engineering reference** for CICFlowMeter-compatible fields

## Out-of-scope use

- Production intrusion detection on real network traffic
- Forensic attribution of real attacks
- Adversarial robustness evaluation (the dataset is not adversarially
  generated)
- Any safety-critical decision

## Reproducibility

Outputs above were produced with `seed = 42`, stratified 70/15/15 split,
on the published sample (`xpertsystems/cyb001-sample`, version 1.0.0,
generated 2026-05-16). The feature pipeline in `feature_engineering.py`
is deterministic and the trained weights in this repo correspond exactly
to the metrics above.

The training script itself is private to XpertSystems. The published
artifacts contain the feature pipeline, model weights, scaler, metadata,
and validation results — sufficient to reproduce inference but not
training.

## Files in this repo

| File | Purpose |
|---|---|
| `model_xgb.json` | XGBoost weights |
| `model_mlp.safetensors` | PyTorch MLP weights |
| `feature_engineering.py` | Feature pipeline (load → engineer → encode) |
| `feature_meta.json` | Feature column order + categorical levels |
| `feature_scaler.json` | MLP input mean/std (XGBoost ignores) |
| `validation_results.json` | Per-class metrics, confusion matrix, architecture |
| `ablation_results.json` | Flow-only vs full feature set comparison |
| `inference_example.ipynb` | End-to-end inference demo notebook |
| `README.md` | This file |

## Contact and full product

The full **CYB001** dataset contains ~685,000 rows across four files
with calibrated A+ benchmark validation. The full XpertSystems.ai
synthetic data catalogue spans 41 SKUs across Cybersecurity, Healthcare,
Insurance & Risk, Oil & Gas, and Materials & Energy.

- 📧 **pradeep@xpertsystems.ai**
- 🌐 **https://xpertsystems.ai**
- 🗂  Dataset: https://huggingface.co/datasets/xpertsystems/cyb001-sample

## Citation

```bibtex
@misc{xpertsystems_cyb001_baseline_2026,
  title  = {CYB001 Baseline Classifier: XGBoost and MLP for Synthetic Network Flow Classification},
  author = {XpertSystems.ai},
  year   = {2026},
  url    = {https://huggingface.co/xpertsystems/cyb001-baseline-classifier},
  note   = {Baseline reference model trained on xpertsystems/cyb001-sample}
}
```