YangYang-Research/web-attack-detection

Web Attack Detection (Hybrid CNN-GRU)

Binary classifier for HTTP request / payload strings. Detects common web attacks (SQLi, XSS, CMDi, and related patterns) for RASP / WAF-style offline scoring.

Paper: Research and Development of a Smart Solution for Runtime Web Application Self-Protection (SOICT '23).

Note: Hub Inference Widgets / Inference Providers do not run this model end-to-end. Input must be a 384-d SentenceTransformer embedding (all-MiniLM-L6-v2), not raw text. Use the code below locally.

Model Details

Field	Value
Developed by	YangYang Research (@noobpk / Le-Thanh Phuc et al.)
Model type	Hybrid CNN + GRU (Keras / TensorFlow)
Language	en (HTTP payloads / web request text)
License	MIT
Finetuned from	Embedding front-end: sentence-transformers/all-MiniLM-L6-v2
Weights file	model.h5

Architecture

1. Input: (batch, 384) — MiniLM sentence embedding
2. Reshape: (batch, 384, 1) for Conv1D
3. CNN branch: Conv1D 32→64→128→256 + MaxPool + GlobalMaxPool → (batch, 256)
4. GRU branch: stacked GRU 32→64→128→256 → (batch, 256)
5. Fusion: element-wise multiply of CNN × GRU outputs
6. Head: Dense 256→128→64→32→1 (sigmoid) — attack probability

Uses

Direct use

• Offline / online scoring of HTTP query strings, body snippets, or path segments
• Research on ML-based RASP / WAF detection
• Complement (not replace) signature WAF / IDS rules

Out of scope

• Standalone production gate without threshold tuning + human review
• Image, audio, or non-sequential tabular classification
• Guaranteed detection of novel / obfuscated zero-days
• Direct use with Hub pipeline("text-classification") (no transformers config.json)

Bias, Risks, and Limitations

• Performance depends on training payload distribution; new frameworks / encodings may drift.
• Overfitting risk on smaller or unbalanced subsets.
• Black-box CNN-GRU: hard to explain individual decisions.
• False positives can block legitimate traffic if threshold is too low.

Recommendations

• Calibrate decision threshold on your traffic.
• Keep signature rules + allowlists alongside the model.
• Retrain / evaluate on domain-specific logs before production.
• Prefer explainability tools (SHAP/LIME on embeddings) for audits.

How to Get Started

pip install -U "keras>=3" tensorflow huggingface_hub sentence-transformers

import os
os.environ["KERAS_BACKEND"] = "tensorflow"

import keras
from huggingface_hub import hf_hub_download
from sentence_transformers import SentenceTransformer

REPO_ID = "YangYang-Research/web-attack-detection"

model_path = hf_hub_download(repo_id=REPO_ID, filename="model.h5")
model = keras.saving.load_model(model_path)
encoder = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")

payload = "1' OR '1'='1 --"
embedding = encoder.encode(payload).reshape(1, 384)
score = float(model.predict(embedding, verbose=0)[0][0])
print(f"attack_probability={score:.4f}")

Keras Hub-style path (weights still loaded from model.h5 via download):

# Equivalent one-liner once weights are local:
# model = keras.saving.load_model("hf://YangYang-Research/web-attack-detection")
# Only works if the repo follows Keras 3 Hub layout (config.json + model.weights.h5).
# This repo currently ships legacy `model.h5` — use hf_hub_download as above.

Training Details

Training data

• Dataset: YangYang-Research/web-attack-detection
  (~626k labeled payloads; ~295k attack / ~331k benign)
• Split: ~70% train / ~30% test

Hyperparameters

Hyperparameter	Value
Optimizer	Adam (lr=0.001)
LR schedule	InverseTimeDecay (steps=1000, rate=0.1)
Batch size	256
Early stopping	patience=3
Dropout	0.1 (branches), 0.3 (fusion)
CV	K-Fold k=5
Loss	binary cross-entropy
Metric	accuracy

Compute

• Google Colab Pro
• Jupyter Notebook

Evaluation

Test split (~30%) of YangYang-Research/web-attack-detection.

Reported metrics: precision, recall, F1, accuracy (see paper / figures below).

Repository layout

File	Role
model.h5	Keras model (architecture + weights)
config.json	Small Hub metadata / architecture summary (not Transformers)
metadata.json	Keras-oriented save metadata
README.md	This model card

Previous Hub layout shipped a ~98 MB config.json that embedded raw weight arrays and claimed library_name: transformers. That file is removed — it broke Hub config parsing (Config file config.json cannot be fetched (too big)).

Citation

@inproceedings{10.1145/3628797.3628901,
  author    = {Le-Thanh, Phuc and Le-Anh, Tuan and Le-Trung, Quan},
  title     = {Research and Development of a Smart Solution for Runtime Web Application Self-Protection},
  year      = {2023},
  isbn      = {9798400708916},
  publisher = {Association for Computing Machinery},
  address   = {New York, NY, USA},
  url       = {https://doi.org/10.1145/3628797.3628901},
  doi       = {10.1145/3628797.3628901},
  booktitle = {Proceedings of the 12th International Symposium on Information and Communication Technology},
  pages     = {304–311},
  numpages  = {8},
  location  = {Ho Chi Minh, Vietnam},
  series    = {SOICT '23}
}

Model Card Authors / Contact

• YangYang Research — org: YangYang-Research
• Original author: noobpk