Preprocessing Pipelines including Block-Matching Convolutional Neural Network for Image Denoising to Robustify Deep...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Cybersecurity

Background:

• Artificial neural networks (ANNs) are crucial for computer vision, including security applications like reidentification.
• ANNs are vulnerable to adversarial attacks, posing risks to surveillance systems.
• Retraining classifiers to counter attacks is computationally expensive.

Purpose of the Study:

• To evaluate the effectiveness of preprocessing defenses against adversarial attacks in computer vision.
• To assess the viability of using image denoising techniques as a defense mechanism.
• To explore defenses that avoid computationally intensive classifier retraining.

Main Methods:

• Tested various preprocessing pipelines against adversarial attacks on a pre-trained neural network.
• Utilized transfer learning to adapt a general architecture for a specific reidentification task.
• Included block-matching convolutional neural network for image denoising as a specific defense.

Main Results:

• Preprocessing defenses demonstrated promising results in mitigating adversarial attacks.
• Image denoising proved effective as an adversarial defense strategy.
• The approach successfully protected a reidentification system in a realistic setting.

Conclusions:

• Preprocessing defenses offer a viable and efficient method to enhance the security of artificial neural networks against adversarial attacks.
• These defenses circumvent the need for expensive classifier retraining.
• The findings support the integration of preprocessing techniques into surveillance pipelines for improved robustness.