Testing Techniques for Critical Deep Learning Applications
summary
Claims concrete results and tools: coverage-guided fuzzing and concolic testing find numerical and adversarial failures, DeepXplore and DeepTest generate domain mutations that break models, and abstract interpretation and solver-based model checking prove some robustness properties. Examples include a 3D-printed turtle misclassified as a rifle, a one-pixel attack flipping a digit, and stop sign patches that evade detection. A system-level pipeline called viateur DSS generates graph-based scenarios and runs Gazebo simulations to expose occlusion and edge-case failures.
AI-written summary · about
chapters (6)
- 0:00Start
- 2:16(Deep) Neural Networks
- 2:57Challenges
- 7:40Example faults: adversarial examples
- 15:27Examples - different lighting conditions
- 19:38Example run (tram example)
