Poster: Camera Tampering Detection for Outdoor IoT Systems

Technical Summary: Camera Tampering Detection for Outdoor IoT Systems

Problem Statement
The proliferation of smart cameras in outdoor surveillance and environmental monitoring has introduced significant reliability challenges. These systems, often deployed in unattended locations, are vulnerable to tampering caused by deliberate vandalism or harsh environmental conditions (e.g., storms). Such tampering manifests as blurred images, rotated lenses, or obstructed views, which degrade monitoring effectiveness. While video-based tampering detection benefits from temporal context provided by continuous frames, detecting tampering in static image capture systems (common in animal detection or periodic monitoring) is more challenging due to the lack of sequential data. Furthermore, existing deep learning solutions often require substantial computational power, making them unsuitable for resource-constrained edge devices like the ESP32 or Raspberry Pi 2.

Methodology
The authors propose and evaluate two distinct approaches for detecting tampered static images: a lightweight rule-based method and a deep learning-based Convolutional Neural Network (CNN) model.

• Hardware and Dataset: Experiments were conducted using a Raspberry Pi 2 Model B with a Camera Module 3. A dataset of 3,600 images was collected over 10 days, categorized into 1,800 normal images, 600 blurred, 600 rotated, and 600 with obstructed lenses.
• Rule-Based Method: This approach establishes a baseline using only normal images captured under various lighting and weather conditions.
  • Feature Matching: It utilizes the ORB (Oriented FAST and Rotated BRIEF) algorithm and Brute Force Matcher (BFMatcher) to compare keypoints between a test image and reference images.
  • Logic Flow: If the number of matches falls below a calibrated threshold, the image is flagged as abnormal. Subsequent checks differentiate between blur (using the Laplacian operator) and empty/obstructed lenses (via pixel intensity standard deviation). If matches exceed the threshold, a homography matrix estimates rotation; angles exceeding 50° are classified as rotated.
  • Requirements: This method requires only 8 reference samples for calibration and no abnormal training data.
• Deep Learning Method: A CNN model implemented with TensorFlow and Keras was designed for binary classification (normal vs. abnormal).
  • Architecture: The model consists of three convolutional layers (32, 64, and 128 filters) with ReLU activation and max-pooling, followed by a fully connected layer (128 units) and a sigmoid output layer.
  • Training: The model was trained on a robust dataset of 2,400 images (1,200 normal, 1,200 abnormal) to ensure generalization across diverse weather and lighting conditions.

Key Results
The performance of both methods was evaluated on a balanced test set of 1,200 images (50% abnormal).

• Accuracy: The CNN model significantly outperformed the rule-based method, achieving an accuracy of 99.75% compared to 90.50% for the rule-based approach. Similarly, the CNN demonstrated superior precision (0.9983 vs. 0.9036) and F1-score (0.9974 vs. 0.9051).
• Resource Efficiency: The rule-based method proved far more lightweight, with a model size of 8.77 KB versus the CNN's 37.8 MB.
• Processing Time: Contrary to expectations regarding deep learning complexity, the CNN model showed a faster average processing time per image (0.02 sec) compared to the rule-based method (0.07 sec) during testing.
• Data Dependency: The rule-based method required only 8 reference samples and zero abnormal samples for training, whereas the CNN required 2,400 labeled samples.

Contributions and Significance
The paper makes three primary contributions to the field of IoT surveillance:

1. Comparative Evaluation: It provides a direct comparison of rule-based and deep learning approaches for static image tampering detection, highlighting the trade-offs between accuracy and resource constraints.
2. Public Dataset: The authors released a publicly available dataset containing normal, blurred, rotated, and obstructed images to support the development and evaluation of future detection algorithms.
3. Practical Deployment Insights: The study identifies that while CNNs offer superior accuracy, the rule-based method remains a viable and necessary alternative for scenarios where:
   • Edge devices have severe memory limitations.
   • Collecting large datasets of abnormal/tampered images is impractical, time-consuming, or costly.
   • Rapid calibration without a prolonged training phase is required.

Conclusion
The authors conclude that the choice of detection method depends heavily on deployment constraints. The CNN model is preferred for high-accuracy requirements where computational resources and data collection are feasible. Conversely, the rule-based method offers a resource-efficient solution for constrained environments where collecting extensive training data is not possible. Future work aims to develop a hybrid system integrating both approaches to balance performance and resource usage.