DenseGAAL: Density-Guided Generative Augmentation for Active Learning in Cybersecurity

---

About the Project

DenseGAAL (Density-Guided Synthetic Augmentation) is a post-query augmentation framework that operates within a standard pool-based active learning loop. After each acquisition round, DenseGAAL:

• Projects the current labeled and unlabeled pools into a representation space induced by the classifier.
• Selects class-dependent anchor points from the labeled pool using a diversity-aware selection procedure that spreads anchors across distinct local regions within each class.
• Retrieves local neighborhoods from the unlabeled pool around each anchor to estimate local support and identify regions where additional augmentation is most needed.
• Allocates augmentation jointly by class imbalance and local density: minority classes receive larger augmentation budgets, while anchors in sparse or isolated regions receive fewer generated samples to avoid unrealistic interpolation.
• Trains local class-conditional generators (TVAE, CTGAN, or RTF) on each anchor neighborhood and samples synthetic examples proportional to the anchor's support score.

This closed-loop, density-aware process adaptively shifts support toward underrepresented regions as the labeled pool evolves—leading to consistent gains in macro-F1 for minority classes without degrading majority-class performance. DenseGAAL is agnostic to the underlying acquisition strategy and classifier architecture, and has been validated across Margin, PowerMargin, Entropy, Density, CLUE, and GALAXY query strategies with Random Forest, XGBoost, MLP, and TabM classifiers.

On cybersecurity datasets (CIC-IDS-2017, CIC-IDS-2018, BODMAS, APIGRAPH, ANDMAL) and non-cybersecurity tabular benchmarks (Shuttle, Cover, Satellite), DenseGAAL consistently outperforms standard active learning and uniform generative augmentation baselines, achieving 3–7 average F1-point gains over BaseAL and 4–10 points over uniform generators at a 1% labeling budget.

---

Motivation

When annotation budgets are tight, active learning often misses rare classes, leading to poor generalization. Existing class-conditional augmentation methods assume enough labeled examples per class to reliably estimate the class distribution—an assumption that routinely fails for minority classes under limited budgets. DenseGAAL addresses this gap by estimating local structure around individual anchor points and selectively expanding underrepresented regions in the feature space using density-guided generative sampling.

---

Built With

• Python 3.x
• Generative modeling: TVAE, CTGAN, RTF (RealTabFormer), TabDDPM
• Classifiers: MLP, Random Forest (RF), XGBoost (XGBC), TabM

---

🛠 Getting Started

Prerequisites

• Python 3.x
• Install dependencies:

  pip install -r requirements.txt

Setup

1. Clone this repo:

   git clone <your-repo-url>
   cd <repo-name>

2. Place your input CSV in source_code/raw_data/.

3. Edit source_code/config.py to define:

   • LABEL_NAME, FEATURE_NAMES, DISCRETE_FEATURES
   • NUM_FEATURE, STANDARDIZE, NUM_CLASS, etc.

4. Example config snippet:

   [config screenshot placeholder]

---

Usage

1. Data Preprocessing

From source_code/, run:

python -m data_pre_process_pipeline \
  --input_csv adult.csv \
  --output_dir adult \
  --label_col income \
  --discrete_to_label

• --discrete_to_label: label-encodes discrete/categorical features. Use ONLY if your discrete features are not numerical.
• output_dir must match the reference name set in config.py.

Produces: data/adult/train.npz, val.npz, test.npz, and label2id.json.

2. Main Experiment Pipeline

Run:

python -m main \
  --al_method <base|DA|DenseGAAL> \
  --al_function <random|entropy|lc|margin|coreset|galaxy|bald|powermargin|clue|diana|eada|upper|lower|density> \
  --classifier <MLP|RF|XGBC> \
  --dataset adult \
  --budget 5 \
  --random_state 42 \
  --generator CTGAN \
  --num_synthetic 3.0 \
  --filter_synthetic \
  --alpha 1.0 \
  --steepness 50.0

Brief argument descriptions

Argument	Description
--al_method	Experiment type
--al_function	Active learning selection strategy
--classifier	Classification model
--dataset	Reference dataset name
--budget	Number of samples per AL round
--random_state	Seed for reproducibility
--generator	Generative model for augmentation
--num_synthetic	Synthetic sample multiplier
--filter_synthetic	Only include filtered synthetic data
--alpha, --steepness	Generation hyperparameters
--decay_power	Controls generation for all classes, (higher decay focuses generation on minority classes)

❗ Note: galaxy and clue AL methods work only with the MLP classifier.

Example Runs

Base Active Learning (no augmentation):

python -m main --al_method base --al_function random --classifier MLP --dataset adult --budget 5

With data augmentation only:

python -m main --al_method DA --al_function random --generator CTGAN --classifier XGBC --dataset adult --budget 5

Full DenseGAAL pipeline:

python -m main --al_method DenseGAAL --al_function random --generator CTGAN --classifier XGBC --dataset adult --budget 5 --filter_synthetic --alpha 1 --steepness 50

---

Project Structure

source_code/
├── active_learning_functions/
├── classifiers/
├── config.py
├── data_pre_process_pipeline.py
├── main.py
├── raw_data/
├── data/
└── results/

---

Results & Benchmarks

Place outputs under results/. DenseGAAL consistently outperforms standard methods, especially in class-imbalanced and distribution-shifted scenarios.