luluzhou1 / dinome

DINoMe

This repository include three submodules:

1. chipyard with our Tiny configurations for NormalCache, ScatterCache, and PhantomCache, defined in chipyard/generators/example/src/main/scala/NormalConfig.scala, chipyard/generators/example/src/main/scala/ScatterConfig.scala, chipyard/generators/example/src/main/scala/PhantomConfig.scala;
2. yosys with a modified SMT2 transformer to generate transition formula and initial state;
3. cryptominisat_DINoMe: Implementation for postcondition construction, and count/ sample <S, S', C, I> tuples;

and two folders:

1. ML: ML related codes for feature engineer and rule generations; (under xgboost)

• Train rules using interference sample and noninterference sample; (train_bits.py)
• Organizing rules by sorting and dropping rules (Fig. 18); (rule.py)

3. process: include bash/python scripts for experiments and software payload for case studies.

• Prerequisite1: You should build yosys and build your verilog code (e.g., make verilog CONFIG=NormalTinyBoomConfigS4W4 in chipyard) to a directory with boom.ys (e.g., process/normal/cache-S4W4).
• Prerequisite2: to composse multi-cycle postcondition, counting and sampling solutions, you should install compose, count, and sampler binaries using https://github.com/DINoMe-Project/cryptominisat
• Under process/normal/cache-S4W4, run yosys boom.ys converts verilog code to SMT2 transition logic and initialization logic for a tiny BOOM with 4-way 4-set cache. This may takes minutes to an hour.
• process/script/cache_observe.sh is one-step script to generate postcondition for cache-based side channel examples, it would generate a directory states with s${cycle}.cnf. This may takes long depending on how many cycles you want. By default, we simplify the CNF per cycle. You can tweak this by updating generate_cnf.sh