Physically Real-time Infrared Attack against Optical Flow Estimation Networks

Optical Flow Estimation Networks (OFENs) are critical for autonomous systems, as they track object motion by analyzing pixel changes between consecutive video frames. These networks typically rely on assumptions like brightness constancy—the idea that an object's brightness remains stable across frames. The authors propose a Real-time Infrared Light Attack (RILA) that exploits these assumptions by using infrared LEDs to inject adversarial perturbations directly into the camera's input stream.

The methodology is divided into two phases. First, a training phase uses a Genetic Algorithm (GA) to generate optimal adversarial examples (AEs) in a controlled physical environment. Because the physical interaction between light and the camera is non-differentiable, the GA evolves light-intensity patterns to maximize the disruption of optical flow estimates. Second, these results are used to train an Adversarial Generative Network (AGN), a lightweight model capable of generating these perturbations in real-time during deployment.

Most adversarial attacks are developed in the digital domain and often fail when transferred to the physical world due to environmental noise, sensor limitations, and fabrication errors. By shifting the paradigm to a physical-train and physical-deploy model, this research demonstrates that OFENs are vulnerable to stealthy, real-time manipulation. Because the attack uses infrared light, it remains imperceptible to human observers, making it a significant security concern for applications like autonomous driving, vehicle tracking, and surveillance, where optical flow is a foundational component for downstream decision-making.