DL-LA: Deep Learning Leakage Assessment

This repository contains source code related to the article DL-LA: Deep Learning Leakage Assessment - A modern roadmap for SCA evaluations and provides download links for obtaining associated trace files.

Short Description of DL-LA

DL-LA is a leakage assessment methodology based on deep learning. Simply put, a neural network is trained to become a classifier in order to distinguish between two groups of side-channel measurements which have been acquired while supplying a cryptographic implementation with one of two distinct fixed inputs (fixed-vs-fixed). In the subsequent validation phase the classification capability of the network is evaluated. In case the classifier succeeds with a (statistically) significantly higher percentage of correct classifications than it could be achieved by a randomly guessing binary classifier, it can be concluded that the side-channel measurements seen by the network during the training phase reveal enough input-dependent information to confidently distinguish the two groups.

Content of the Repository

In this repository we share simple Python (and C++) scripts that allow you to perform deep learning leakage assessment on your own side-channel data and partially reproduce the results reported in the DL-LA paper. Please note, that all given scripts are reduced to their basic functionality and have been kept short and simple on purpose. To use the scripts in this repository a working installation of TensorFlow and Keras is required. The scripts have been tested using tensorflow-gpu 2.1.0 and keras-gpu version 2.4.3. For reproducibility of (a part of) the experimental results presented in the paper we have hosted the underlying leakage traces for two case studies, namely CS3 and CS5 (see DL-LA). Those trace files can be downloaded here:

• CS3: Serialized PRESENT implementation, randomized clock (heavily misaligned), measured on SAKURA-G FPGA board: LINK
• CS5: Serialized PRESENT threshold implementation, bad trigger (slightly misaligned), measured on SAKURA-G FPGA board: LINK

Getting Started

The simplest way of getting started is to download the side-channel traces linked above (and in the respective subfolders). However, you may also start right away with your own SCA data. In any case, just follow the steps below to get everything up and running:

• Step 1: edit/check traces.yml and make sure that all trace sets which should be analyzed are correctly defined
• Step 2: edit/check extract_mean_and_std_deviation.py and make sure that all trace sets which should be analyzed are listed under "names"
• Step 3: execute the extract_mean_and_std_deviation.py script - this may take some time depending on the size of your SCA data but only needs to be executed once at the beginning or whenever the trace sets or their definitions have changed
• Step 4: choose one of the three DL-LA scripts and adapt, if necessary, the name of the trace set and the training/validation parameters:
  1. DL-LA_MLP_SA_inputs: DL-LA using the standard MLP network with final sensitivity analysis based on the network inputs
  2. DL-LA_MLP_SA_first_layer_weights: DL-LA using the standard MLP network with final sensitivity analysis based on the first layer weights
  3. DL-LA_CNN_SA_inputs: DL-LA using the standard CNN network with final sensitivity analysis based on the network inputs
• Step 5: execute the script that was chosen and adapted in Step 4
• Step 6: find the validation accuracy results in val_acc.log (ASCII) and the sensitivity analysis results in sensi.dat (BINARY)
• Step 7: you may want to use the C++ multi-precision log probability calculator (Boost C++ library required) in order to convert the validation accuracy values and the size of the validation set into log probabilities
• Step 8: Repeat from Step 4 for further analyses on the same trace sets

Sensitivity Analysis Troubleshooting

In our experience it can be difficult at times to obtain nice and smooth Sensitivity Analysis (SA) results (even when the training and validation accuracy look promising). Therefore, we want to provide a couple of hints here. First of all the sensitivity analysis performed in our scripts is related to the training set and the training loss only. If you want to improve the sensitivity analysis results increase the number of training traces (and decrease the number of validation traces to save time). Furthermore, we have implemented two different variants of the sensitivity analysis for the MLP network. The first is based on the network inputs (this is also applicable to the CNN but for SA the MLP should be favored) and the second one is based on the first layer weights. Try both and compare the results. In our experience, the inputs-based sensitivity is often superior. Finally, the number of epochs the network is trained plays a very significant role in the sensitivity analysis result. Make sure to try different values for the number of epochs (e.g., 5, 10, 25, 50, 100) if computationally affordable. If not, values around 10 usually give decent results for SA.

Contact and Support

Please contact Thorben Moos (thorben.moos@rub.de) if you have any questions, comments or if you found a bug that should be corrected.

Licensing

Please see LICENSE for licensing instructions.

Publication

DL-LA: Deep Learning Leakage Assessment - A modern roadmap for SCA evaluations.