Falsification-Driven Robust Training
Using adversarial falsification to expand certified robustness radii by 48%.
- Period
- August–November 2025
- Kind
- course
- Stack
- CARLA, VerifAI, AutoLiRPA, PyTorch
- Guide
- Prof. Supratik Chakraborty, IIT Bombay
- Built a pipeline coupling the CARLA simulator with VerifAI to synthesise adversarial examples, then used them to train models that are provably more robust.
- Computed certified L2 robustness radii with Alpha-CROWN via AutoLiRPA, which sidestepped the VRAM and latency bottlenecks that made standard Branch-and-Bound infeasible at this scale.
- Falsification-driven active learning expanded certified L2 robustness radii by 48% over five training cycles, showing markedly better sample efficiency than random sampling.