Adversarial Autoencoder (AdvAE) and MLP Training Script

This script provides functionality to train and evaluate a Multi-Layer Perceptron (MLP) and an Adversarial Autoencoder (AdvAE) for adversarial instance detection. It is highly configurable via command-line arguments for flexible usage.

The paper is Adversarial Detection and Correction by Matching Prediction Distributions

The script is implemented based on Adversarial Auto-Encoder. An example is provided that works on [CIFAR10] dataset(https://docs.seldon.io/projects/alibi-detect/en/stable/examples/ad_ae_cifar10.html).

---

Features

• Train an MLP classifier.
• Load pre-trained MLP weights.
• Train an AdvAE detector.
• Evaluate the AdvAE detector on adversarial data.

---

Requirements

• Python 3.7+
• Required Python libraries:
  • numpy
  • tensorflow
  • sklearn
  • alibi-detect

Install dependencies:

conda create -n my_ae_detector python=3.9
conda activate my_ae_detector
pip install alibi-detect[tensorflow]
pip install tensorflow[and-cuda]==2.14.1

2. Example Commands

Train MLP and AdvAE

python advAE_script.py --mlp --train --data_path "/data/qgong/data/AE/CICDDoS2019/Data/cicddos" --mlp_model_path "mlp_classifier.weights.h5" --model_save_path "adv_ae_detector_with_pretrained_classifier"

Load Pre-trained MLP and Train AdvAE

python advAE_script.py --train --data_path "/data/qgong/data/AE/CICDDoS2019/Data/cicddos" --mlp_model_path "mlp_classifier.weights.h5" --model_save_path "adv_ae_detector_with_pretrained_classifier"

Train MLP, Train AdvAE, and Perform Prediction

python advAE_script.py --mlp --train --eval --data_path "/data/qgong/data/AE/CICDDoS2019/Data/cicddos" --mlp_model_path "mlp_classifier.weights.h5" --model_save_path "adv_ae_detector_with_pretrained_classifier" --threshold_perc 90.0

Load Pre-trained MLP and AdvAE, and Perform Prediction

python advAE_script.py --eval --data_path "/data/qgong/data/AE/CICDDoS2019/Data/cicddos" --mlp_model_path "mlp_classifier.weights.h5" --model_save_path "adv_ae_detector_with_pretrained_classifier" --threshold_perc 90.0

---

Data Requirements

The dataset should be structured as follows:

• X_train.npy: Features for training.
• y_train.npy: Labels for training.
• X_test.npy: Features for testing.
• y_test.npy: Labels for testing.
• Adv/Adv_X_test.npy: Features for adversarial examples.

Ensure all .npy files are present in the directory specified by the --data_path argument.

---

Workflow

Step 1: Train MLP

If --mlp is provided:

1. Train the MLP classifier using the training dataset (X_train, y_train).
2. Save the trained weights to the file specified by --mlp_model_path.

Step 2: Train AdvAE

If --train is provided:

1. Use the MLP classifier (trained or loaded) as a pre-trained model.
2. Train the AdvAE detector using the training dataset.
3. Save the trained AdvAE model to the path specified by --model_save_path.

Step 3: Evaluate AdvAE

If --eval is provided:

1. Load the trained AdvAE model from the path specified by --model_save_path.
2. Infer the detection threshold using the training dataset.
3. Predict adversarial instances on the adversarial dataset (Adv/Adv_X_test.npy).

---

Outputs

1. MLP Model Weights: Saved at the path specified by --mlp_model_path (default: classifier.weights.h5).

2. AdvAE Detector: Saved at the path specified by --model_save_path (default: adv_ae_detector_with_pretrained_classifier).

3. Evaluation Metrics:

   • Instance scores for adversarial detection.
   • Adversarial prediction results.

---