FNet: A Two-Stream Model for Detecting Adversarial Attacks against 5G-Based Deep Learning Services

Table 4

The detailed two-stream network architecture for CIFAR-10. Conv (d, k, s) denotes the convolutional layer with d as dimension, k as kernel size, and s as stride.

RGB stream	SRM stream
Conv (64, 3, 1)	Conv (64, 3, 1)
BatchNorm layer, Tanh	ABSLayer
Avg pooling	BatchNorm layer, Tanh
Conv (128, 3, 1)	Avg pooling
BatchNorm layer, Tanh	Conv (128, 3, 1)
Avg pooling	BatchNorm layer, Tanh
Conv (256, 3, 1)	Avg pooling
BatchNorm layer, ReLU	Conv (256, 3, 1)
Conv (256, 3, 1)	BatchNorm layer, ReLU
BatchNorm layer, ReLU	Conv (256, 3, 1)
MAX pooling	BatchNorm layer, ReLU
	MAX pooling
Full connected 4096, ReLU, dropout (0.5)
Full connected 4096, ReLU, dropout (0.5)
Softmax 2