🎯 Adversarial Attacks on Video Anomaly Detection Models in AIoT

This repository contains code and resources for generating and evaluating adversarial attacks on deep learning-based video anomaly detection models. The project is focused on the intersection of AIoT (Artificial Intelligence of Things) and Adversarial Machine Learning, specifically targeting real-world surveillance scenarios such as UCF Crime.

---

📌 Project Overview

Recent research has shown that deep neural networks, especially those used in video anomaly detection, are vulnerable to adversarial attacks — small perturbations in the input that lead to misclassification. This repository implements:

• One Pixel Attack
• Multi-Pixel Attack
• A novel Multi-Pixel Deception (MPD) attack: combining the power of One Pixel and Pixel attacks
• Preprocessing pipelines for video anomaly datasets
• Evaluation scripts and visualizations

---

🗂 Directory Structure

📦 root
├── adversarial_samples      # Generated adversarial examples
├── notebooks                # Jupyter notebooks for attack demos
├── preprocess               # Scripts for trimming, resizing, and augmenting videos
├── LICENSE                  # MIT License
├── README.md                # You're here!
├── moondream_reqs.txt       # Optional dependencies for vision-language models
├── requirements.txt         # Python dependencies

---

🚀 Getting Started

1. Clone the repository

git clone https://github.com/qaixerabbas/adv_attacks_vad.git
cd adv_attacks_vad

2. Install Dependencies

pip install -r requirements.txt

Optional: For vision-language filtering (Moondream (currently adopted) or you can use TinyLLaVA), install from moondream_reqs.txt.

---

📽 Dataset

This code is tested on the UCF Crime dataset, a large-scale real-world surveillance dataset. You may need to request access to the dataset separately from UCF Crime Dataset.

---

🧠 Models

• ✅ ResNet-18
• ✅ EfficientNet-B0
• ✅ MobileNet-v3 Small
• (Plug-and-play architecture: easily extendable to more CNN models)

---

🔐 Attack Goals

• Mislead anomaly detection models with high success rate
• Maintain imperceptibility of adversarial perturbations
• Evaluate robustness across diverse architectures

---

📝 Citation

If you use this code in your research, please cite:

@article{hina2025adversarial,
  title={Adversarial attacks on artificial Intelligence of Things-based operational technologies in theme parks},
  author={Hina, Sadaf and Abbas, Qaiser and Ahmed, Kashan},
  journal={Internet of Things},
  pages={101654},
  year={2025},
  publisher={Elsevier}
}

---

⚖ License

This project is licensed under the MIT License.

---

🤝 Contributing

Contributions are welcome! Feel free to fork this repository, raise issues, and submit pull requests.

---

📬 Contact

For any inquiries or collaborations, reach out to:
Qaiser Abbas – [mqaiser617@gmail.com]