ludwigw/trm-cifar10-pgd

TRM CIFAR-10 PGD

TRM trained with PGD-7 adversarial training on CIFAR-10 at ε=8/255.

Training Configuration

• Method: PGD-7 adversarial training
• ε: 8/255 (L∞ on [0,1] RGB pixels)
• Input: 32×32 RGB (3072 dimensions)
• Architecture: TRM-MLP adapted for CIFAR-10 (105.8M params)
• Epochs: 50

Verification Results (β-CROWN, 512 samples)

ε (L∞)	Verified	vs IBP
0.001	94%	+16%
0.002	90%	+39%
0.004	80%	+70%
0.006	67%	+66%
0.008	58%	+58%

Winner on CIFAR-10: PGD dominates complex datasets with 94% at ε=0.001. IBP completely fails.

Key Insight

Dataset complexity determines training method effectiveness:

• MNIST (784 dim, simple): IBP wins (78% vs 63%)
• CIFAR-10 (3072 dim, complex): PGD wins (94% vs 78%)

Usage

import torch
from veriphi.models import TinyRecursiveMLP

model = TinyRecursiveMLP(x_dim=3072, y_dim=512, z_dim=512, hidden=1024,
                         num_classes=10, H_cycles=2, L_cycles=2)
model.load_state_dict(torch.load("trm-cifar10-pgd.pt"))
model.eval()

# CIFAR-10 input: flatten 32x32x3 to 3072
x = torch.randn(1, 3072)
logits = model(x)

Citation

@article{deshmukh2026veriphi,
  title={Veriphi: Attack-Guided Neural Network Verification with Dataset-Dependent Training Methods},
  author={Deshmukh, Pratik and Savin, Vasili and Arya, Kartik},
  year={2026}
}

Paper: arXiv:XXXX.XXXXX | Code: GitHub