Invisible CMOS Camera Dazzling for Conducting Adversarial Attacks on Deep Neural Networks.

Area of Science:

• Computer Vision
• Cybersecurity
• Optics

Background:

• Deep neural networks (DNNs) excel in performance but are susceptible to adversarial attacks.
• Existing physical adversarial attacks are often visually detectable by humans.
• Vulnerabilities in physical systems pose significant cybersecurity risks.

Purpose of the Study:

• To introduce a novel, invisible optical-based physical adversarial attack targeting CMOS cameras.
• To analyze the conditions necessary for an attack to be imperceptible to the human eye yet effective against DNNs.
• To investigate the relationship between light source parameters and attack efficacy.

Main Methods:

• Designing a specific light pulse sequence for optical attack.
• Utilizing the camera's shutter mechanism to spatially transform the light pulse within the image.
• Analyzing photopic conditions for invisibility and image disruption.
• Evaluating the attack's success rate against DNNs under varying light source duty cycles.

Main Results:

• Demonstrated an invisible optical-based physical adversarial attack on CMOS cameras.
• Identified optimal photopic conditions for the attack to remain undetected by humans.
• Showcased the ability to deceive DNNs using the proposed method.
• Quantified the trade-off between attack success and concealment based on light source duty cycle.

Conclusions:

• The proposed invisible optical attack is a viable method for deceiving DNNs in the physical world.
• Controlling the light source duty cycle is crucial for balancing attack effectiveness and stealth.
• This research highlights new physical vulnerabilities in camera-based AI systems.