posit-nn-fault-injection

Posit NN fault injection. Code developed during the "Architetture dei sistemi di elaborazione" special project.

Project structure

This project is structured as follows:

Project tree

SP5/
├─ data/
│  ├─ dataset_name_1/
│  │  ├─ model_name_1/
│  │  ├─ model_name_2/
│  ├─ dataset_name_2/
├─ models/
├─ res/
│  ├─ dataset_name_1/
│  │  ├─ model_name_1/
│  │  ├─ model_name_2/
│  ├─ dataset_name_2/
├─ src/
├─ utils/
├─ main.py

Folder content

data/ -> Containes the weights obtained in the training phase. Each dataset and model has its own folder
models/ -> Contains the custom models used
res/ -> Containes the results of the FI campaign. Each dataset and model has its own folder
src/ -> Contains the files that manages the FI campaigns
utils/ -> Folder with utility files (parser, parameters getters, ...)
main.py -> Main file to perform FI campaigns and inference

Setup

Prerequisite:

• Python 3.6 (strictly required)
• protobuf 3.19.6 (strictly required)

pip install protobuf==3.19.6

In order to use this FI framework the following python packages are required:

• numpy 1.15.2
• softposit
• numpy-posit (a modified version of numpy supporting posit data type)
• tensorflow-posit (a modified version of tensorflow supporting posit data type)
• scipy

You can install these packages with pip, using the following commands (the creation of a virtual env is recommended):

pip install requests numpy==1.15.2 softposit

pip install numpy-posit

pip install https://s3-ap-southeast-1.amazonaws.com/posit-speedgo/tensorflow_posit-1.11.0.0.0.1.dev1-cp36-cp36m-linux_x86_64.whl

pip install scipy

The order of the commands is important.

For more detailed instructions you can check Deep PeNSieve installation guide.

Usage

To run the programm it is necessary set some parameters from command line. Required values are:

--type or -t                #Numeric format:  posit8/posit16/posit32/float32        
--network-name or -n        #Network to use:  convnet
--data-set or -d            #Input dataset:   CIFAR10     
--bit-len or -b             #Number of bit:   8/16/32/32
                            #It is critical that this value be consistent with the numeric format

Other commands may be optional because they use some default values, but they can be modified using:

--batch-size                #Batch size dimension  
--size or -s                #Test set size
--force-n                   #Force to a specific number of injections
--seed                      #Seed to random value
--net-level                 #Index in the net where apply faults
--low-index                 #To set minimum index to start to apply fault on the weight
--high-index                #To set maximum index to finish to apply fault on the weight
--name-output               #To set name of the output file

To watch more details and its default value you can run the command:

--help or -h

Output

Program produce in output a csv file, if file already exists, it writes to the bottom. Otherwise it creates new one, which name is <type>_injection.csv.

Output is organized into:

• Fault_id: identifier fault
• Layer_index: location in the net where fault is applied
• Tensor_index: index where apply the fault
• Bit_index: weight bit subject to injection
• Accuracy: net accuracy with corrupt weight
• Golden_accuracy: net accuracy without corrupt weight
• Difference: difference between accuracy and golden_accuracy
• Top_5: net top 5 accuracy
• Weight_difference: difference between golden_weight and corrupted_weight

Credits

The ML tasks performed in this framework rely on: Deep PeNSieve.