adds simulation, attack, and metrics related to optical PUFs

Co-Authored-By: Adomas Baliuka <[email protected]>

CITATION.cff

@@ -20,4 +20,6 @@ authors:
     given-names: Niklas
   - family-names: Mursi
     given-names: Khalid T.
+  - family-names: Baliuka
+    given-names: Adomas
 license: GPL 3.0

README.md

@@ -39,7 +39,8 @@ To refer to pypuf, please use DOI `10.5281/zenodo.3901410`.
 pypuf is published [via Zenodo](https://zenodo.org/badge/latestdoi/87066421).
 Please cite this work as
 
-> Nils Wisiol, Christoph Gräbnitz, Christopher Mühl, Benjamin Zengin, Tudor Soroceanu, Niklas Pirnay, & Khalid T. Mursi.
+> Nils Wisiol, Christoph Gräbnitz, Christopher Mühl, Benjamin Zengin, Tudor Soroceanu, Niklas Pirnay, Khalid T. Mursi,
+> & Adomas Baliuka.
 > pypuf: Cryptanalysis of Physically Unclonable Functions (Version 2, June 2021). Zenodo.
 > https://doi.org/10.5281/zenodo.3901410
 
@@ -53,7 +54,8 @@ or use the following BibTeX:
                   Benjamin Zengin and
                   Tudor Soroceanu and
                   Niklas Pirnay and
-                  Khalid T. Mursi},
+                  Khalid T. Mursi and
+                  Adomas Baliuka},
   title        = {{pypuf: Cryptanalysis of Physically Unclonable
                    Functions}},
   year         = 2021,

docs/appendix/bibliography.rst

@@ -36,6 +36,7 @@ Bibliography
 .. [NSJM19] Nguyen, P. H. et al. The Interpose PUF: Secure PUF Design against State-of-the-art Machine Learning Attacks.
     IACR Transactions on Cryptographic Hardware and Embedded Systems 243–290 (2019) doi:10.13154/tches.v2019.i4.243-290.
 .. [ODon14] O’Donnell, R. Analysis of Boolean Functions. (Cambridge University Press, 2014).
+.. [RHUWDFJ13] Rührmair, U. et al. Optical PUFs Reloaded. https://eprint.iacr.org/2013/215 (2013).
 .. [RSSD10] Rührmair, U. et al. Modeling Attacks on Physical Unclonable Functions. in Proceedings of the 17th ACM
     Conference on Computer and Communications Security 237–249 (ACM, 2010). doi:10.1145/1866307.1866335.
 .. [SD07] Suh, G. E. & Devadas, S. Physical Unclonable Functions for Device Authentication and Secret Key Generation.

docs/attacks/linear_regression.rst

@@ -0,0 +1,116 @@
+Linear Regression
+=================
+
+Linear Regression fits a linear function on given data. The resulting linear function, also called map, is guaranteed
+to be optimal with respect to the total squared error, i.e. the sum of the squared differences of actual value and
+predicted value.
+
+Linear Regression has many applications, in pypuf, it can be used to model :doc:`../simulation/optical` and
+:doc:`../simulation/arbiter_puf`.
+
+
+Arbiter PUF Reliability Side-Channel Attack [DV13]_
+---------------------------------------------------
+
+For Arbiter PUFs, the reliability for any given challenge :math:`c` has a close relationship with the difference in
+delay for the top and bottom line. When modeling the Arbiter PUF response as
+
+.. math::
+    r = \text{sgn}\left[ D_\text{noise} + \langle w, x \rangle \right],
+
+where :math:`x` is the feature vector corresponding to the challenge :math:`c` and :math:`w \in \mathbb{R}^n` are the
+weights describing the Arbiter PUF, and :math:`D_\text{noise}` is chosen from a Gaussian distribution with zero mean
+and variance :math:`\sigma_\text{noise}^2` to model the noise, then we can conclude that
+
+.. math::
+    \text{E}[r(x)] = \text{erf}\left( \frac{\langle w, x \rangle}{\sqrt{2}\sigma_\text{noise}} \right).
+
+Hence, the delay difference :math:`\langle w, x \rangle` can be approximated based on an approximation of
+:math:`\text{E[r(x)]}`, which can be easily obtained by an attacker. It gives
+
+.. math::
+    \langle w, x \rangle = \sqrt{2}\sigma_\text{noise} \cdot \text{erf}^{-1} \text{E}[r(x)].
+
+This approximation works well even when :math:`\text{E}[r(x)]` is approximated based on only on few responses, e.g. 3
+(see below).
+
+To demonstrate the attack, we initialize an Arbiter PUF simulation with noisiness chosen such that the reliability
+will be about 91% *on average*:
+
+>>> import pypuf.simulation, pypuf.io, pypuf.attack, pypuf.metrics
+>>> puf = pypuf.simulation.ArbiterPUF(n=64, noisiness=.25, seed=3)
+>>> pypuf.metrics.reliability(puf, seed=3).mean()
+0.908...
+
+We then create a CRP set using the *average* value of responses to 500 challenges, based on 5 measurements:
+
+>>> challenges = pypuf.io.random_inputs(n=puf.challenge_length, N=500, seed=2)
+>>> responses_mean = puf.r_eval(5, challenges).mean(axis=-1)
+>>> crps = pypuf.io.ChallengeResponseSet(challenges, responses_mean)
+
+Based on these approximated values ``responses_mean`` of the linear function :math:`\langle w, x \rangle`, we use
+linear regression to find a linear mapping with small error to fit the data. Note that we use the ``transform_atf``
+function to compute the feature vector :math:`x` from the challenges :math:`c`, as the mapping is linear in :math:`x`
+(but not in :math:`c`).
+
+>>> attack = pypuf.attack.LeastSquaresRegression(crps, feature_map=lambda cs: pypuf.simulation.ArbiterPUF.transform_atf(cs, k=1)[:, 0, :])
+>>> model = attack.fit()
+
+The linear map ``model`` will predict the delay difference of a given challenge. To obtain the predicted PUF response,
+this prediction needs to be thresholded to either -1 or 1:
+
+>>> model.postprocessing = model.postprocessing_threshold
+
+To measure the resulting model accuracy, we use :meth:`pypuf.metrics.similarity`:
+
+>>> pypuf.metrics.similarity(puf, model, seed=4)
+array([0.902])
+
+
+Modeling Attack on Integrated Optical PUFs [RHUWDFJ13]_
+-------------------------------------------------------
+
+The behavior of an integrated optical PUF token can be understood as a linear map
+:math:`T \in \mathbb{C}^{n \times m}` of the given challenge, where the value of :math:`T` are determined by the given
+PUF token, and :math:`n` is number of challenge pixels, and :math:`m` the number of response pixels.
+The speckle pattern of the PUF is a measurement of the intensity of its electromagnetic field at the output, hence the
+intensity at a given response pixel :math:`r_i` for a given challenge :math:`c` can be written as
+
+.. math::
+    r_i = \left| c \cdot T \right|^2.
+
+pypuf ships a basic simulator for the responses of :doc:`../simulation/optical`, on whose data a modeling attack
+can be demonstrated. We first initialize a simulation and collect challenge-response pairs:
+
+>>> puf = pypuf.simulation.IntegratedOpticalPUF(n=64, m=25, seed=1)
+>>> crps = pypuf.io.ChallengeResponseSet.from_simulation(puf, N=1000, seed=2)
+
+Then, we fit a linear map on the data contained in ``crps``. Note that the simulation returns *intensity* values rather
+than *field* values. We thus need to account for quadratic terms using an appropriate
+:meth:`feature map <pypuf.attack.LeastSquaresRegression.feature_map_optical_pufs_reloaded_improved>`.
+
+>>> attack = pypuf.attack.LeastSquaresRegression(crps, feature_map=pypuf.attack.LeastSquaresRegression.feature_map_optical_pufs_reloaded_improved)
+>>> model = attack.fit()
+
+The success of the attack can be visually inspected or quantified by the :doc:`/metrics/correlation` of the response
+pixels:
+
+>>> crps_test = pypuf.io.ChallengeResponseSet.from_simulation(puf, N=1000, seed=3)
+>>> pypuf.metrics.correlation(model, crps_test).mean()
+0.69...
+
+Note that the correlation can differ when additionally, post-processing of the responses is performed, e.g. by
+thresholding the values such that half the values give -1 and the other half 1:
+
+>>> import numpy as np
+>>> threshold = lambda r: np.sign(r - np.quantile(r.flatten(), .5))
+>>> pypuf.metrics.correlation(model, crps_test, postprocessing=threshold).mean()
+0.41...
+
+
+API
+---
+
+.. autoclass::
+    pypuf.attack.LeastSquaresRegression
+    :members: __init__, fit, model, feature_map_optical_pufs_reloaded, feature_map_optical_pufs_reloaded_improved

docs/index.rst

@@ -17,6 +17,7 @@ Some functionality is only implemented in the
    simulation/arbiter_puf
    simulation/delay
    simulation/bistable
+   simulation/optical
    simulation/base
 
 .. toctree::
@@ -36,6 +37,7 @@ Some functionality is only implemented in the
    metrics/reliability
    metrics/uniqueness
    metrics/distance
+   metrics/correlation
    Fourier-Analysis (PUFMeter) <metrics/fourier>
 
 .. toctree::
@@ -47,6 +49,7 @@ Some functionality is only implemented in the
    Logistic Regression <attacks/lr>
    Multilayer Perceptron <attacks/mlp>
    LMN (PUFMeter) <attacks/lmn>
+   attacks/linear_regression
 
 .. toctree::
    :maxdepth: 2
@@ -105,8 +108,8 @@ pypuf is published `via Zenodo <https://zenodo.org/badge/latestdoi/87066421>`_.
 Please cite this work as
 
     Nils Wisiol, Christoph Gräbnitz, Christopher Mühl, Benjamin Zengin, Tudor Soroceanu,
-    Niklas Pirnay, & Khalid T. Mursi.
-    pypuf: Cryptanalysis of Physically Unclonable Functions (Version v2, March 2021). Zenodo.
+    Niklas Pirnay, Khalid T. Mursi, & Adomas Baliuka.
+    pypuf: Cryptanalysis of Physically Unclonable Functions (Version v2, August 2021). Zenodo.
     https://doi.org/10.5281/zenodo.3901410
 
 or use the following BibTeX::
@@ -118,7 +121,8 @@ or use the following BibTeX::
                       Benjamin Zengin and
                       Tudor Soroceanu and
                       Niklas Pirnay and
-                      Khalid T. Mursi},
+                      Khalid T. Mursi and
+                      Adomas Baliuka},
       title        = {{pypuf: Cryptanalysis of Physically Unclonable
                        Functions}},
       year         = 2021,

docs/metrics/correlation.rst

@@ -0,0 +1,26 @@
+Correlation
+-----------
+
+The correlation metric is useful to judge the prediction accuracy when responses are non-binary, e.g. when studying
+:doc:`/simulation/optical`.
+
+>>> import pypuf.simulation, pypuf.io, pypuf.attack, pypuf.metrics
+>>> puf = pypuf.simulation.IntegratedOpticalPUF(n=64, m=25, seed=1)
+>>> crps = pypuf.io.ChallengeResponseSet.from_simulation(puf, N=1000, seed=2)
+>>> crps_test = pypuf.io.ChallengeResponseSet.from_simulation(puf, N=1000, seed=3)
+>>> feature_map = pypuf.attack.LeastSquaresRegression.feature_map_optical_pufs_reloaded_improved
+>>> model = pypuf.attack.LeastSquaresRegression(crps, feature_map=feature_map).fit()
+>>> pypuf.metrics.correlation(model, crps_test).mean()
+0.69...
+
+Note that the correlation can differ when additionally, post-processing of the responses is performed, e.g. by
+thresholding the values such that half the values give -1 and the other half 1:
+
+>>> import numpy as np
+>>> threshold = lambda r: np.sign(r - np.quantile(r.flatten(), .5))
+>>> pypuf.metrics.correlation(model, crps_test, postprocessing=threshold).mean()
+0.41...
+
+.. automodule:: pypuf.metrics
+    :members: correlation, correlation_data
+    :noindex:

docs/simulation/optical.rst

@@ -0,0 +1,8 @@
+Integrated Optical PUFs
+=======================
+
+pypuf ships a very basic simulation of integrated optical PUFs [RHUWDFJ13]_.
+
+.. autoclass::
+    pypuf.simulation.IntegratedOpticalPUF
+    :members: __init__, eval

docs/simulation/overview.rst

@@ -15,6 +15,7 @@ pypuf currently features simulation of the following strong PUFs:
    Lightweight Secure PUF [MKP08]_, Permutation PUF [WBMS19]_, and Interpose PUF [NSJM19]_.
 2. :doc:`Feed-Forward Arbiter PUFs <arbiter_puf>` and XORs thereof [GLCDD04]_.
 3. :doc:`PUF designs based on bistable rings <bistable>` [CCLSR11]_ [XRHB15]_.
+4. :doc:`Integrated Optical PUFs <optical>` [RHUWDFJ13]_.
 
 Technicalities
 --------------

pypuf/attack/__init__.py

@@ -3,3 +3,5 @@
 from .mlp2021 import MLPAttack2021
 
 from .fourier import LMNAttack
+
+from .linear_regression import LeastSquaresRegression

pypuf/attack/linear_regression.py

@@ -0,0 +1,104 @@
+from typing import Callable, Optional
+
+import numpy as np
+
+from .base import OfflineAttack
+from ..io import ChallengeResponseSet
+from ..simulation.base import Simulation
+
+
+class LinearMapSimulation(Simulation):
+
+    @staticmethod
+    def postprocessing_id(responses: np.ndarray) -> np.ndarray:
+        return responses
+
+    @staticmethod
+    def postprocessing_threshold(responses: np.ndarray) -> np.ndarray:
+        return np.sign(responses)
+
+    def __init__(self, linear_map: np.ndarray, challenge_length: int,
+                 feature_map: Optional[Callable[[np.ndarray], np.ndarray]] = None,
+                 postprocessing: Optional[Callable[[np.ndarray], np.ndarray]] = None) -> None:
+        super().__init__()
+        self.map = linear_map
+        self._challenge_length = challenge_length
+        self.feature_map = feature_map or (lambda x: x)
+        self.postprocessing = postprocessing or self.postprocessing_id
+
+    @property
+    def challenge_length(self) -> int:
+        return self._challenge_length
+
+    @property
+    def response_length(self) -> int:
+        return self.map.shape[1]
+
+    def eval(self, challenges: np.ndarray) -> np.ndarray:
+        return self.postprocessing(self.feature_map(challenges) @ self.map)
+
+
+class LeastSquaresRegression(OfflineAttack):
+
+    @staticmethod
+    def feature_map_linear(challenges: np.ndarray) -> np.ndarray:
+        return challenges
+
+    @staticmethod
+    def feature_map_optical_pufs_reloaded(challenges: np.ndarray) -> np.ndarray:
+        """
+        Computes features of an optical PUF token using all ordered pairs of challenge bits [RHUWDFJ13]_.
+        An optical system may be linear in these features.
+
+        .. note::
+            This representation is redundant since it treats ordered paris of challenge bits are distinct.
+            Actually, only unordered pairs of bits should be treated as distinct. For applications, use
+            the function :meth:`feature_map_optical_pufs_reloaded_improved
+            <pypuf.attack.LeastSquaresRegression.feature_map_optical_pufs_reloaded_improved>`,
+            which achieves the same with half the number of features.
+
+        :param challenges: array of shape :math:`(N, n)` representing challenges to the optical PUF.
+        :return: array of shape :math:`(N, n^2)`, which, for each challenge, contains the flattened dyadic product of
+            the challenge with itself.
+        """
+        beta = np.einsum("...i,...j->...ij", challenges, challenges)
+        return beta.reshape(beta.shape[:-2] + (-1,))
+
+    @staticmethod
+    def feature_map_optical_pufs_reloaded_improved(challenges: np.ndarray) -> np.ndarray:
+        r"""
+        Computes features of an optical PUF token using all unordered pairs of challenge bits [RHUWDFJ13]_.
+        An optical system may be linear in these features.
+
+        :param challenges: 2d array of shape :math:`(N, n)` representing `N` challenges of length :math:`n`.
+        :return: array of shape :math:`(N, \frac{n \cdot (n + 1)}{2})`. The result `return[i]` consists of all products
+            of unordered pairs taken from `challenges[i]`, which has shape `(N,)`.
+
+        >>> import numpy as np
+        >>> import pypuf.attack
+        >>> challenges = np.array([[2, 3, 5], [1, 0, 1]])  # non-binary numbers for illustration only.
+        >>> pypuf.attack.LeastSquaresRegression.feature_map_optical_pufs_reloaded_improved(challenges)
+        array([[ 4,  6, 10,  9, 15, 25],
+               [ 1,  0,  1,  0,  0,  1]])
+        """
+        n = challenges.shape[1]
+        idx = np.triu_indices(n)
+        return np.einsum("...i,...j->...ij", challenges, challenges)[:, idx[0], idx[1]]
+
+    def __init__(self, crps: ChallengeResponseSet,
+                 feature_map: Optional[Callable[[np.ndarray], np.ndarray]] = None) -> None:
+        super().__init__(crps)
+        self.crps = crps
+        self.feature_map = feature_map or (lambda x: x)
+        self._model = None
+
+    def fit(self) -> Simulation:
+        features = self.feature_map(self.crps.challenges)
+        # TODO warn if more than one measurement
+        linear_map = np.linalg.pinv(features) @ self.crps.responses[:, :, 0]
+        self._model = LinearMapSimulation(linear_map, self.crps.challenge_length, self.feature_map)
+        return self.model
+
+    @property
+    def model(self) -> Simulation:
+        return self._model

pypuf/metrics/__init__.py

@@ -1,3 +1,3 @@
 from .common import reliability_data, reliability, uniqueness_data, uniqueness, similarity_data, accuracy, similarity, \
-    bias, bias_data
+    bias, bias_data, correlation, correlation_data
 from .fourier import total_influence, noise_sensitivity, influence