Compressive imaging for defending deep neural networks from adversarial attacks.

Area of Science:

• Computer Vision
• Machine Learning Security
• Signal Processing

Background:

• Deep neural networks (DNNs) excel in performance but are susceptible to adversarial perturbations.
• Adversarial attacks pose a significant threat to the reliability of DNNs.

Purpose of the Study:

• To introduce a novel defense mechanism against adversarial attacks on DNNs.
• To leverage compressive sensing (CS) for both defense and image encoding.

Main Methods:

• Employing compressive sensing (CS) to defend DNNs against adversarial perturbations.
• Utilizing CS for image encoding to prevent counterattacks.
• Conducting computer simulations and optical experiments with a CS single pixel camera.

Main Results:

• Demonstrated the effectiveness of CS in defending DNNs against adversarial attacks.
• Showcased the dual role of CS in defense and image encoding.
• Validated the approach through simulations and real-world optical experiments.

Conclusions:

• Compressive sensing offers a robust solution for defending DNNs against adversarial attacks.
• The proposed CS-based method provides enhanced security by preventing counterattacks.
• This approach shows promise for secure object classification in adversarial conditions.