As deep neural networks become integral to safety-critical computer vision applications, ensuring their reliability requires both rigorous physical stress-testing and efficient defensive design. This presentation addresses the dual imperatives of vision safety, bridging the gap between dynamic real-world threat vectors and proactive architectural hardening.
We begin by exposing fundamental physical vulnerabilities using RILA, a real-time infrared attack framework targeting Optical Flow Estimation Networks (OFENs). Operating imperceptibly to the human eye via 840nm infrared LEDs, RILA employs a two-stage optimization paradigm—combining offline Genetic Algorithms with an Adversarial Generative Network (AGN)—to generate dynamic illumination vectors in under 30 ms. By disrupting brightness constancy, RILA successfully forces state-of-the-art architectures like RAFT and PWC-Net into motion tracking failures across moving targets and varying distances.
To counter such systemic brittleness, we transition to TRNAS, an ultra-fast, training-free Neural Architecture Search (NAS) framework that fundamentally redesigns vision backbones for intrinsic resilience. By replacing costly adversarial training with a novel zero-cost proxy (R-Score) and a Multi-Objective Selection strategy, TRNAS evaluates candidates at initialization based on linear activation expressivity and feature map consistency. The framework navigates a search space of neural network  architectures in just 0.02 GPU days, discovering models that achieve top-tier robust accuracy across CIFAR-10, CIFAR-100, and ImageNet benchmarks. Together, these complementary breakthroughs reveal that dense early-layer topology and physical-domain evaluation are essential pillars for building truly dependable computer vision systems.