Defensive ML + Maya

This repository integrates Defensive ML techniques into Maya.
Most of the framework remain the same, except that we replace Maya's Gaussian Sinusoid planner with a Defensive ML defender. You can do either of followings:

• Use the pre-trained defender to protect the system from power
• Train a new defender by collecting the power traces from your system

Requirements

1. Install PARSEC 3.0 benchmark and modify PARSECDIR in Scripts/Launch.sh, line 184.
2. Follow instructions of Maya-README.md to set up the Maya framework in your system.
3. Install PyTorch. Running the ML-related code requires a CUDA-capable GPU.
4. Install libtorch API and set TORCHDIR in Makefile to the directory where libtorch is installed.
5. make will build the system.

Defending the system with the pre-trained defender

1. Running ./defender_wrapper.sh 0 [samplecount] will turn on the Maya defense with Defensive ML, and collect the PARSEC benchmark traces while the defense is active. Recommended samplecount: 100 (total 1000 traces for 10 benchmarks). It will take 5~6 hours to collect 1000 traces.
2. Running python MayaDataset.py --victimdir defender_logs will emulate the attack on the power side-channel, giving the CNN accuracy.

Training a new defender

1. Run ./wrapper.sh 0 [samplecount] to collect unprotected traces for PARSEC benchmark runs. Recommended samplecount: 400 (total 4000 traces). It will take a full day to collect 4000 traces.
2. Run python DefensiveML.py to train the ML defender.
3. Change name of the output pth file cp parsec_shaper_64_[...].pth best_shaper_64.pth
4. Follow shaperToScript.ipynb to get the jit-traced cpuscript_parsec_shaper_64.pt. Copy it under Controller/ directory.
   cp cpuscript_parsec_shaper_64.pt Controller/ssvFast3_Shaper_64.pt
   Or, you can use the provided Controller/ssvFast3_Shaper_64.pt
5. Run ./defender_wrapper.sh 0 [samplecount] to collect the traces with the defender activated.
6. Run ./maya_wrapper.sh 0 [samplecount] to collect the traces with Maya defense activated.
7. Run MayaDataset.py --victimdir [defender_logs|maya_logs] to evaluate the defense results for ML defense (defender_logs) or Maya defense (maya_logs)