Challenging AI+Camera Systems with Physical Adversarial Attacks

The integration of cameras and artificial intelligence (AI) has revolutionized various applications, yet the robustness of these systems against physical adversarial attacks remains underexplored. Existing adversarial strategies typically target single imaging devices, resulting in limited efficacy and stability across different camera systems. Here, we present a camera-agnostic physical adversarial attack framework designed to challenge AI+camera synergies. Our approach introduces a novel adversarial optimization framework that consists of an attack module and a defense module. The attack module optimizes perturbations to maximize their effectiveness, while the defense module adjusts the parameters of a camera ISP proxy to minimize the attack's impact. These modules interact in an adversarial game, enhancing the stability of attacks across multiple camera types. Through extensive experiments, we demonstrate the efficacy of our method in misleading AI-driven tasks across various hardware configurations, including two distinct cameras and four smartphones. This work advances the field of physical adversarial attacks by providing robust, real-world applicable techniques for undermining AI+camera systems.