ResNet-20 With Multi-Channel Striding and Slot Management

This example demonstrates the ResNet-20 FHE pipeline with two main features:

1. Multi-channel striding – downsampling handled through combined convolution + shortcut blocks.

2. Dynamic slot switching – layers run at different slot sizes by clearing and reloading rotation/bootstrapping keys.

Overview

The workflow includes:

1. FHE Context and Key Generation

   • Initialize FHEController and ANNController.

   • Generate bootstrapping and rotation keys per layer with different slot values.

2. Data Preparation

   • Load convolution kernels, shortcut kernels, and fully connected weights.

   • Encode weights and biases for each block.

3. Multi-Channel Striding

   • Strided residual blocks are processed using secure_double_optimized_convolution_multi_channels() for efficient conv + shortcut handling.

4. Slot Management

   • Before each major block (layer group or FC), contexts are cleared and reloaded with the correct slot size. This is adopted after every downsampling block.

5. Inference Loop

   • Encrypt CIFAR-10 images.

   • Apply convolution and residual blocks with striding + shortcut connections.

   • Dynamically adjust slot counts across layers.

   • Perform global average pooling and final FC layer for classification.

Key Functions

• convolution_block() – Load and encode kernels, run optimized convolution.

• shortcut_convolution_block() – Encode shortcut kernels for downsampling.

• double_shortcut_convolution_block() – Perform strided convolution + shortcut in parallel (multi-channel).

• resnet_block() – Residual block with bootstrap, striding, shortcut, and secure ReLU.

• fc_layer_block() – Encode and apply the final fully connected layer.

Data Processing Functions

• load_weights() – Load convolution kernels.

• load_fc_weights() – Load FC or shortcut kernels.

• load_bias() – Load bias vectors.

• Encoding handled by optimized FHE encoders (per conv/shortcut/FC).

Block Functions

• convolution_block(...) – Single convolution + bias.

• shortcut_convolution_block(...) – Downsample shortcut branch.

• double_shortcut_convolution_block(...) – Conv + shortcut with striding.

• resnet_block(...) – Full residual block with ReLU and bootstrap.

• fc_layer_block(...) – Final FC classification layer.

Full Example Source

ResNet20MultiChannelStriding.cpp