Soft body pose-invariant evasion attacks against deep learning human detection

Evasion attacks on deep neural networks (DNN) use manipulated data to let targets evade detection and/or classification across a wide range of DNNs. Most existing evasion attacks focus on planar images (e.g., photo, satellite imaging) and ignore the distortion of evasion in practical attacks (e.g., object rotation, deformation). Here, we build evasion patterns for soft-body human stakeholders, where patterns are designed to take into account body rotation, fabric stretch, printability, and lighting variations. We show that these are effective and robust to different human poses. This poses a significant threat to safety of autonomous vehicles and adversarial training should consider this new area.