Efficient Attack Detection with Multi-Latency Neural Models on Heterogeneous Network Devices (ICNP 2023)

To get started, follow these steps:

1. Install the required dependencies:

pip install -r requirements.txt  
# The latest mmt package is required  
git clone https://github.com/makerlin1/MMT.git --depth=1  
cd MMT  
python install build  
python install install  

2. Make sure to install the C++ version of MNN on both your server and target network device.

Prepare your dataset, which should be organized as a dictionary with keys "train", "valid", and "test" for input data (X) and output labels (y). The dataset should be loaded using torch.utils.data.Dataset (check the botdataset.py for more details.). Next, run the train.sh script for NAS training. Before executing this script, remember to set the --data_path in the script as the path of your dataset (e.g., we set it as dataset.pkl).

sh train.sh  

Note: Feel free to customize the hyperparameters in the YAML files within the NSGA directory to better suit your specific task.

First, create a new directory named ops_list and run the get_ops_list.py program (located in the root directory). This script generates all the basic operators (Ops) of the supernet in the ops_list directory.

python get_ops_list.py  

To measure the real latency of Ops on your target device, run the following code:

from mmt.meter import meter_ops  
meter_ops("ops_list")  

Next, create a new directory named models in the root directory. The device latency re-sorting is implemented in the resort.py script. Run the script using the following command:

python resort.py  

The final output will be similar to the example below, displaying the suggested models with their accuracy and latency:

Model name                                            Acc    Latency  
------------------------------------------------  -------  ---------  
./models/soteria_acc1_96.322434_latency_0.122467  96.3224   0.122467  
./models/soteria_acc1_98.415320_latency_0.193567  98.4153   0.193567  
./models/soteria_acc1_99.070213_latency_0.302434  99.0702   0.302434  
./models/soteria_acc1_98.510031_latency_0.264667  98.51     0.264667  
...                                              ...      ...  

For a detailed explanation of the detection process, please refer to the README.md file in the Detection folder. To perform mixed inference, you need to set the model candidates in line 43 of the MNN_process.cpp file as follows:

string g_modelList[MAX_MODEL_COUNT] = {     // These models need to be sorted from lowest to highest accuracy  
       "../../models/soteria_acc1_96.322434_latency_0.122467.mnn",  
       "../../models/soteria_acc1_98.415320_latency_0.193567.mnn",  
       "../../models/soteria_acc1_98.510031_latency_0.264667.mnn",  
        "../../models/soteria_acc1_99.070213_latency_0.302434.mnn"  
};  

Sort the models from the lowest to the highest accuracy according to the output from the device latency re-sorting step.