โข
Abstract:
โฆ
We propose a lightweight neural network classifier that detects unstable frames in physics-based animation (MPM) simulations. To ensure the reliability of predictions under pixel-level perturbations, we apply formal verification using the nnenum tool. Our model is trained on a custom dataset of successful and failed simulations generated by a differentiable MPM-based morphing framework. We also apply adversarial training methods (FGSM, PGD) and evaluate certified robustness using CROWN-IBP.
โฆ
Dataset: 6000+ frames (Stable/Unstable) extracted from custom morphing simulationsโฆ
Goal: Real-time detection and formal robustness certification of unstable frames in graphics rendering Results
โฆ
Classification Accuracy:โช
PGD-trained model achieved 99.88% test accuracy
โช
Regular & FGSM-trained models: 99.25%
โฆ
Robustness Verification (nnenum):โช
No models passed global verification at even small perturbation (ฯต = 0.02)
โช
PGD-trained model showed localized certifiable robustness on high-confidence samples (up to ฯต = 0.07)
โฆ
Model Size:โช
CNN2 achieved state-of-the-art performance with <1M parameters, making it verifiable and efficient
โฆ
Dataset:โช
3694 stable frames / 2308 unstable frames
โช
Generated from custom differentiable MPM morphing simulations
