Smart-Card SCA ML Pipeline
Budget: $750 – $1,500 USD
I am building an end-to-end side-channel analysis workflow that ingests two-channel traces (instantaneous power and I/O) captured with a ChipWhisperer Husky and drives them through classic Correlation Power Analysis before moving into a PyTorch-based machine-learning stage.
My single objective is to recover cryptographic material—specifically RSA and 3DES keys—directly from those traces. You will start with the unlabelled samples attached here and, once we agree on the approach, I will hand over roughly 10 000 additional datasets for full-scale training and validation.
What I need from you is a reproducible pipeline that:
• pre-processes and aligns the raw Husky files,
• runs a solid CPA baseline to identify points of interest,
• designs and trains a PyTorch model (CNN, RNN, or a hybrid you justify) that improves key-recovery success over the CPA baseline,
• outputs recovered keys together with confidence metrics, and
• produces clear documentation so I can rerun or extend the analysis later.
Acceptance criteria will be met when the delivered code can reliably extract the target keys on a withheld validation set and all steps are fully scripted for headless execution on a Linux workstation equipped with a modern GPU.
My single objective is to recover cryptographic material—specifically RSA and 3DES keys—directly from those traces. You will start with the unlabelled samples attached here and, once we agree on the approach, I will hand over roughly 10 000 additional datasets for full-scale training and validation.
What I need from you is a reproducible pipeline that:
• pre-processes and aligns the raw Husky files,
• runs a solid CPA baseline to identify points of interest,
• designs and trains a PyTorch model (CNN, RNN, or a hybrid you justify) that improves key-recovery success over the CPA baseline,
• outputs recovered keys together with confidence metrics, and
• produces clear documentation so I can rerun or extend the analysis later.
Acceptance criteria will be met when the delivered code can reliably extract the target keys on a withheld validation set and all steps are fully scripted for headless execution on a Linux workstation equipped with a modern GPU.