Camera Tampering Detection

Surveillance systems are vulnerable to physical tampering—unauthorized alterations of the camera’s viewpoint, defocusing, or masking. Traditional detection methods rely on short-term temporal buffers, which are easily fooled by slow, gradual tampering. This project proposes a Generative Reference Model to establish a “ground truth” of normal operating conditions.

Technical Approach

• Generative Adversarial Network (GAN): We trained a GAN to learn the probability density function of images from a specific camera under normal conditions. This allows the system to “sample” a reference image that represents the expected scene at any time.
• Siamese Network: To compare the live feed against the generated reference, we employed a Siamese architecture. This network transforms both images into a high-dimensional feature space, maximizing the distance between “normal” and “tampered” states.
• Deep Learned Features: By moving away from handcrafted edges or SIFT features, the model achieves higher robustness against environmental noise like rain, snow, or moving shadows.

Key Results

The proposed method significantly reduces false alarms caused by dynamic backgrounds while maintaining a near-perfect detection rate for abrupt or malicious camera movements.