CalyPSO

This repository contains accompanying code for "CalyPSO: An Enhanced Search Optimization based Framework to Model Delay Based PUFs" (to be) presented at CHES 2024.

Setup

Clone the repository to the destination directory (say CALYPSO_CLONE_DIR) and setup packages through pip3 install -r requirements.txt.

Running calypso

1. Navigate through cd CALYPSO_CLONE_DIR/pypuf

2. Execute python3 calypso.py --target-degree 1 --challenge-num 50000 --cut-length 2 --proc 1 --population 500 to launch calypso as per relevant parameterization:

   • --target-degree: The value of k in a k-XOR PUF. Example: --target-degree 1 mounts the attack on a APUF
   • --challenge-num: The number of challenge response pairs to use in training.
   • --cut-length: The number of delay stages to change for each generation in the evolutionary algorithm. For a k-XOR PUF, valid range is from 2 to (64*k - 1)
   • --proc: The number of processors to spread the workload on
   • --population: Initial size of the population

3. Additional optional flags:

   • -landscape-evolution: Whether to include landscape evolution or not. By default, it is disabled. Upon enabling this switch, in 10 generations, 25% of the CRPs are evolved to smooth out the problem landscape over a period of time.
   • -aeomebic-reproduction: Whether to enable aeomebic reproduction or not. By default, it is disabled.

Sample command to enable both: python3 calypso.py --target-degree 1 --challenge-num 50000 --cut-length 64 --proc 1 --population 500 -landscape-evolution -aeomebic-reproduction

Running calypso++

Similar to running calypso, except now, calypso++.py must be used. Consequently, --target-degree k represents an attack on k LP-PUF using a cross-architectural model of a k-XOR PUF. For example, to model 1 LP-PUF using a population of APUFs, execute python3 calypso++.py --target-degree 1 --challenge-num 50000 --cut-length 2 --proc 1 --population 500 -landscape-evolution -aeomebic-reproduction

Jaccard distance

Both calypso.py and calypso++.py rely upon accuracy and bias metrics to construct their fitness functions. However, Jaccard distance could also be used as an alternative. We provide a sample in calypso_jaccard.py, which uses Jaccard distance to construct its fitness function. Rest all semantics of calypso_jaccard.py are similar to calypso.py

Hardware data

We have also uploaded the hardware data we have collected and used to test CalyPSO/CalyPSO++ on real devices. Under CALYPSO_CLONE_DIR/datasets, there are two subsections: CALYPSO_CLONE_DIR/datasets/in-house (holding the collected challenge-response tuples for in-house generated hardware) and CALYPSO_CLONE_DIR/datasets/online (holding data from https://pypuf.readthedocs.io/en/latest/data/datasets.html)

The user can specify hardware data of a new PUF of their choice, with the restriction that both the challenge file and the response file must have only a single subdatabase. To elaborate, calypso_hardware.py shall load the challenge data as challenge_data = np.load(args.challenge_file) and then extract the first subdatabase as challenge_data.files[0] to load the actual challenges. The challenges can be present as a bitstring sampled from {0, 1}^n (where n is the challenge length; n=64 in our experiments). We internally convert such challenges to {-1, 1}^n.

We give some more examples here:

BR-PUF

We have also added a dedicated version of calypso (named calypso_hardware.py) to aid in testing out the hardware data. A sample run command: python3 calypso_hardware.py --target-degree 4 --cut-length 64 --challenge-num 100000 --proc 1 --population 500 -aeomebic-reproduction --challenge-file $CALYPSO_CLONE_DIR/datasets/in-house/BRPUF/Challenge/chal_64_nChal_200000_bi.npz --response-file $CALYPSO_CLONE_DIR/datasets/in-house/BRPUF/GoldenResponses/respG_BRPUF_64_NChal_200000_5_meas_Br_10_all.npz. This loads the challenge-response data related to BR-PUF and models it (cross-architecturally) using a 4-XOR PUF. There are 3 main changes wrt. the simulation attacks mentioned before:

1. We have removed landscape evolution, since in some cases, we will not have the expansive dataset to consider only a subset of it for training.

2. --challenge-file flag takes the absolute path of the .npz file containing the challenge set

3. --response-file flag takes the absolute path of the .npz file containing the response set

(11-11) iPUF

Some datasets are larger than what GitHub allows in its normal storage. To avoid using Git Large File Storage, we ship such data as zipped. One example is the challenge set of (11-11) iPUF. To run CalyPSO on it, perform the following steps:

1. unzip $CALYPSO_CLONE_DIR/datasets/in-house/11_11_iPUF/chal_64_300000_bi.npz.zip to unzip the challenge set (say to $CALYPSO_CLONE_DIR/datasets/in-house/11_11_iPUF/)
2. Run calypso as python3 calypso_hardware.py --target-degree 11 --cut-length 64 --challenge-num 100000 --proc 1 --population 500 -aeomebic-reproduction --challenge-file $CALYPSO_CLONE_DIR/datasets/in-house/11_11_iPUF/chal_64_300000_bi.npz --response-file $CALYPSO_CLONE_DIR/datasets/in-house/11_11_iPUF/respG_11_11_iPUF_64_NChal_300000_5_means_Br_54.npz to model the (11-11) iPUF cross-architecturally using a 11-XOR PUF.

10XOR-APUF

1. unzip $CALYPSO_CLONE_DIR/datasets/in-house/10XOR-APUF/chal_64_300000_bi.npz.zip to $CALYPSO_CLONE_DIR/datasets/in-house/10XOR-APUF.
2. Run calypso as python3 calypso_hardware.py --target-degree 10 --cut-length 64 --challenge-num 100000 --proc 1 --population 500 -aeomebic-reproduction --challenge-file $CALYPSO_CLONE_DIR/datasets/in-house/10XOR-APUF/chal_64_300000_bi.npz --response-file $CALYPSO_CLONE_DIR/datasets/in-house/10XOR-APUF/respG_10_XORPUF_64_NChal_300000_5_meas_Br_98.npz to model the 10XOR-APUF cross-architecturally using a 10-XOR PUF.