Skip to main content

Table 1 Model architecture for MNIST (left) and CIFAR10 (right) networks

From: Deep neural rejection against adversarial examples

Layer typeDimensionLayer typeDimension
Conv. + ReLU32 filters (3×3)Conv. + Batch Norm. + ReLU64 filters (3×3)
Conv. + ReLU32 filters (3×3)Conv. + Batch Norm. + ReLU64 filters (3×3)
Max Pooling2×2Max Pooling + Dropout (p=0.1)2×2
Conv. + ReLU64 filters (3×3)Conv. + Batch Norm. + ReLU128 filters (3×3)
Conv. + ReLU64 filters (3×3)Conv. + Batch Norm. + ReLU128 filters (3×3)
Max pooling2×2Max Pooling + Dropout (p=0.2)2×2
Fully connected + ReLU200 unitsConv. + Batch Norm. + ReLU256 filters (3×3)
Fully connected + ReLU200 unitsConv. + Batch Norm. + ReLU256 filters (3×3)
Softmax10 unitsMax Pooling + Dropout (p=0.3)2×2
  Conv. + Batch Norm. + ReLU512 filters (3×3)
  Max Pooling + Dropout (p=0.4)2×2
  Fully Connected512 units
  Softmax10 units