Professional Documents
Culture Documents
TonElecDev NO EXP
TonElecDev NO EXP
4, APRIL 2022
Abstract — Previous works have shown excellent PUFs [1]. Memristive PUFs utilize spatial device-to-device
prospects for implementing strong physical unclonable variations, e.g., in I –V nonlinearity (NL) [3]–[5], in an array
functions (PUFs) with memristive crossbar circuits. Here we of memory cells to generate random responses. Promising
first propose two techniques for boosting the robustness of
such PUFs to machine learning (ML) attacks. The general results were also reported for reliability and statistical proper-
idea behind both proposals is to maximize the contribution ties of generated keys, as well as physical performance. The
of each crosspoint device to the PUF output to make main goal of this article is to improve further the robustness of
the response less predictable. Specifically, we present the memristive strong PUFs based on an architecture presented
results for choosing an optimal ratio of selected rows and in [3], [5], and [6]. It should be noted that our focus is not
columns, and investigate in detail the improvements in
robustness due to the balancing of device conductances on memristive PUF circuits, which utilize (small/large signal)
in the crossbar array. The effectiveness of the proposed switching characteristics of memory cells or write speed vari-
algorithm for conductance balancing is confirmed by ations [7]–[10], because of the apparent severe flaws in their
modeling the response of a two-sided PUF based on a 20 × designs. The reliability is a major concern in such PUFs due to
20 crossbar memristive circuit with a multilayer perceptron the limited switching endurance of memristors and their large
network. Second, we explore some open questions
which require in-depth analysis. Specifically, we quantify cycle-to-cycle variations. Additionally, the operation for most
the effect of device nonlinearity and device analog- of those PUFs is based on the intrinsic statistical properties
tunability. We show that nonlinear, analog memristive of a single memory cell. For example, a challenge is used
PUFs outperform the PUFs that have either linear or digital to select a single device, and the response is simply read-out
devices. Finally, we explore the effect of stuck-at-fault digital state. The implication is, first, the challenge-response
devices (nonideal yield) on PUFs uniformity. Indeed,
by modeling this hardware imperfection, we show that the pair (CRP) space is a linear function of array size and hence
proposed algorithm results in a more-robust PUF. cannot be increased too much, and second, the response is a
simple function of the input. As a result, it seems that most
Index Terms — Hardware security, machine learning (ML)
modeling attack, memristor, physical unclonable func- such PUFs are weak, and their advantages compared to pure
tion (PUF), resistive random access memory (ReRAM), CMOS weak PUFs are still not clear.
robustness. In addition to the PUF robustness improvement, some of the
open questions in designing the strong memristive PUFs are
I. I NTRODUCTION explored in this article. For example, the effect of device NL
0018-9383 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Massachusetts Amherst. Downloaded on August 24,2022 at 14:59:20 UTC from IEEE Xplore. Restrictions apply.
LARIMIAN et al.: IMPROVING MACHINE LEARNING ATTACK RESILIENCY VIA CONDUCTANCE BALANCING IN MEMRISTIVE STRONG PUFs 1817
Authorized licensed use limited to: University of Massachusetts Amherst. Downloaded on August 24,2022 at 14:59:20 UTC from IEEE Xplore. Restrictions apply.
1818 IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 69, NO. 4, APRIL 2022
Fig. 3. (a) Simplified equivalent model of M × M crossbar circuit when read voltage Vread is applied to sM selected rows, the output currents are
read from sM virtually grounded columns, while all the remaining lines are floated. (b) and (c) Effect of different column selection ratio in 32 ×
32 crossbars (b) on NIST test suite and (c) on MLP prediction accuracy.
the tuning voltage is sampled from a Gaussian distribution with output, corresponding to the PUF output, while the number
an average of 33.3 μA (corresponding to ∼8.3 μS conductance of layers/neurons in the hidden layer(s) were varied in the
at 0.25 V) and a specific normalized standard deviation σ . simulations. A rectified linear (sigmoid) activation function
After unique NL and current at 0.25 V have been assigned for was used for the hidden (output) layer neurons. The MLP
each device, the constants a and b (and hence complete unique classifier was trained and validated using the conventional
static I –V characteristics for all devices in the crossbar circuit) backpropagation method on 80% of the simulated CRPs. The
are derived from (1). It is noted that the described approach trained network is then used to predict PUF response on the
for choosing currents crudely corresponds to the uncertainty remaining, mutually exclusive 20% of the CRPs.
in the tuning process for configurable PUFs [3], [5]. As it
should be clear from the discussion below, the absolute values V. PUF O PTIMIZATION
for the currents are not important for this particular study due This section first describes two proposed techniques for
to the focus on the functional characteristics of the PUFs and improving robustness against machine learning attacks. The
the assumption of the ideal peripheral circuits and negligible common rationale for both techniques is that PUF robustness
IR drops on the electrodes. The absolute values are only used is increased when all crosspoint devices in the crossbar equally
for a convenience of relating the results to the representative contribute to the output currents. The PUF response, in this
device technology, while a more relevant parameters for this case, would be a nontrivial function of the input, which
study–NL and σ are explicitly specified, whenever appropriate. depends on the unique I –V characteristics of all devices in
the crossbar array. Furthermore, this section explores how the
IV. E VALUATION M ETRICS unique features of memristors namely device NL and analog
To assess the performance of the PUF, we consider three tunability excel PUF security metrics. Moreover, this section
main metrics, e.g., UF, NIST, and predictability which are studies the effect of crossbar size on PUF predictability.
widely discussed in the literature [3], [12], [13]. It is noted
that it is not feasible to measure bit error rate (BER) in this A. Optimal Selection Ratio
article because the temperature variation and conductance drift This section proposes a technique to improve PUF robust-
models are not available. Additionally, note that uniqueness ness against machine learning attacks by maximizing the
is more relevant in evaluating experimental data. Furthermore, contribution of the current of all devices in IL and IR . The
note that diffuseness is a weaker PUF metric as it stays near to requirement for the balanced contribution can be simplified
ideal value (50%) even if other metrics show low performance. to having currents via selected devices similar to those via
To perform predictability analyses, machine learning models (type A) half-selected devices, given that the output current
are considered as they are currently the most effective attack is the sum of these two parts. The circuit parameters for
form for strong PUFs [6]. Memristive crossbar PUF has a having similar currents can be found from the approximate
nonlinear input–output relationship, a vast CRP space, and equivalent circuit of the crossbar array [Fig. 3(a)], which
a time-independent output response. As a result, multilayer is derived assuming negligible line resistance and similar
perceptron (MLP) is chosen as an attack over logistic regres- static I –V characteristics of all crosspoint devices. Using
sion (used for linear-separable data), support vector machine the approximate equivalent circuit, the selected current and
(runs very slow for huge data), and recurrent neural network the leakage current can be written as s 2 M 2 asinh(bVS ) and
and long short-term memory (both use history of data). The s(1 − s)M 2 a sin h(bV© ), respectively.
studied MLP network consists of 2M inputs so that the Our preliminary analysis for the considered average NL
challenge can be directly applied to the MLP input, and one shows that selection ratios n/M = 0.25 and m/M = 0.2, i.e.,
Authorized licensed use limited to: University of Massachusetts Amherst. Downloaded on August 24,2022 at 14:59:20 UTC from IEEE Xplore. Restrictions apply.
LARIMIAN et al.: IMPROVING MACHINE LEARNING ATTACK RESILIENCY VIA CONDUCTANCE BALANCING IN MEMRISTIVE STRONG PUFs 1819
Authorized licensed use limited to: University of Massachusetts Amherst. Downloaded on August 24,2022 at 14:59:20 UTC from IEEE Xplore. Restrictions apply.
1820 IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 69, NO. 4, APRIL 2022
Fig. 7. MLP prediction accuracy of the two-sided PUF (M = 20, n = 5, m = 4) response as a function of (a) number of CRPs used in training
40-100-1 network, (b) number of hidden layer neurons in two-layer network, and (c) number of hidden layers in 40-100-. . . -100-1 network. There
different scenarios for crossbar conductances scenarios were simulated: σ = 2% (blue circle symbols); σ = 25% without applying algorithm (red
cross symbols); and σ = 25% with applying algorithm (green square symbols). In panel (b) and (c) studies, 50K CRPs were used for training MLP
network. The error bars represent standard deviation for the five simulated PUF instances, each with different device I–V characteristics.
average conductance of all devices in the simulated instance of in the former study case, likely due to the limited number of
the crossbar array. The first/second term in the cost function is CRPs. Surprisingly, the accuracy is almost independent of the
a sum of squared differences between the conductance of the number of hidden layers for the latter. Finally, just like for
row/column and the global average value. The cost function the first experiment, the algorithm allows reducing prediction
optimizes PUF for both one-sided and two-sided architectures accuracy for the PUFs with σ = 25% device conductance
and is independent of the number of selected columns. distribution to that of naive one with σ = 2%, which confirms
A simulated annealing approach was implemented so that a the algorithm’s effectiveness.
move is always accepted if the cost function is reduced, while It should be noted that the crossbar (either naive or balanced
it is accepted with a certain probability, determined by the one) is secure against side-channel attacks because generating
change in the cost and the current annealing temperature, even IL and IR do not reveal any information about their relative
if the cost is increased. The annealing parameters are chosen values [13]–[17]. To study this claim, the power profile of
such that most of the memristors are swapped multiple times. 5K CRPs of ten 20 × 20 PUF instances have been collected.
Fig. 6(a)–(c) shows an example of applying an algorithm for Statistics showed an ideal fraction of 50% of the times when
a 10 × 10 crossbar array. The sum of the conductances across response = one consumed more power than response = 0.
rows and columns has significant dispersion for the initial, ran-
dom distribution of conductances [Fig. 6(a)], while these sums C. Device NL and Analog Tunability
become very close to each other after applying the algorithm The memristive PUFs have been widely studied in literature
[Fig. 6(b)]. Fig. 6(c) shows how the value of the cost function before [1], [3], [18], [19], though the detailed studies of
reduces after each iteration. If the algorithm temperatures the impact of memristor NL and analog tunability on PUF
and annealing rate are not carefully selected, the number of security metrics are missing. Such earlier work showed that the
iterations might not be big enough to result in a well-balanced former device feature leads to nonlinear PUF operation which
crossbar. The algorithm effectiveness is investigated for differ- makes the modeling attacks almost impossible. The latter
ent scenarios of the machine learning attacks (Fig. 7). In the device feature leads to tunable PUF which means the devices
first study, the prediction accuracy of the MLP network was can be custom-tuned for a specific goal in the configuration
studied as a function of the number of CRPs used in training phase. As we showed in the previous section, and in particular
for three cases of the crossbar conductances [Fig. 7(a)]. For Fig. 3(a) results, the analog-tunability of memristors could
a smaller number of CRPs, the accuracy is close to the ideal be utilized to implement crossbar with balanced, uniform
50% when the device conductance distribution in the crossbar conductance characteristics, thus making output random and
array is very tight. The accuracy is more than 60%, on average, independent of the input.
for the naïve (random) mapping with σ = 25%, though the To study the effect of the device NL, 50K CRPs of five
application of the algorithm allows reducing it to the ideal nonlinear and linear 20 × 20 PUF instances are simulated.
value. As expected, increasing the number of CRPs makes Moreover, to further study the effect of device analog tun-
machine learning attacks more effective, though the prediction ability, 50K CRPs of five analog and digital 20 × 20 PUF
accuracy seems to saturate, which might be related to the instances are simulated. Specifically, in analog PUFs, the
limited capacity of the used MLP network. target conductances are chosen as explained in Section III.
The impact of the MLP capacity is further investigated A similar operation of digital PUFs is assumed (Section II),
by increasing the number of hidden layer neurons for the while their target conductances are chosen from two Gaussian
two-layer network [Fig. 7(b)] and increasing the number of distributions that are centered on ON- and OFF-conductance
hidden layers while fixing the number of hidden layer neurons values (mean values of 8.3 and 1 μS, and standard devia-
[Fig. 7(c)]. The accuracy improves initially and then saturates tions of 2% and 2%, respectively). Based on the simulation
Authorized licensed use limited to: University of Massachusetts Amherst. Downloaded on August 24,2022 at 14:59:20 UTC from IEEE Xplore. Restrictions apply.
LARIMIAN et al.: IMPROVING MACHINE LEARNING ATTACK RESILIENCY VIA CONDUCTANCE BALANCING IN MEMRISTIVE STRONG PUFs 1821
D. Crossbar Size
Another design variable that affects PUF predictability is
crossbar size. When the crossbar size increases, the PUF Fig. 10. UF values for 20 × 20 PUF as a function of yield for the
complexity increases, which results in a less predictable PUF cases (a) without and (b) with applied balancing algorithm. UF values
behavior. To study this claim, 1M CRPs for 20 × 20, 30 × 30, as a function of PUF size for the cases (c) without and (d) with the
applied balancing algorithm. In all panels, 5K CRPs were used for the five
40 × 40, and 50 × 50 PUF instances are collected, while their simulated PUF instances, each with different device I–V characteristics.
responses are modeled with 2M × 100 × 1 MLP networks.
The realistic values of μG = 8.3 μS with 25% variations in the
target conductance distribution as well as balancing algorithm In this case, the column with the stuck-at fault device may
are considered for all cases. As shown in Fig. 9, when the size become dominant, resulting in a bias in the output which
increases, the MLP accuracy reduces. makes PUF unreliable. To simulate the effect of yield on PUF
reliability, we collected 5K CRPs of five PUF instances for
VI. H ARDWARE I MPERFECTIONS a variety of resistive random access memory (ReRAM)-based
Previous studies explored how IR-drop ([20]) affect PUF PUF sizes which have different percentages of stuck-at ON
robustness. As explained in [20], when the wire resistance (without loss of generality) devices. The realistic values of
of interconnects is nonzero ideal, it causes IR drop along μG = 8.3 μS with 2% (without balancing algorithm) and 25%
interconnects. In fact, the devices that are closer/further than (with balancing algorithm) variations in the target conductance
the voltage source will have a lower/higher IR drop along distribution are considered for all cases.
the interconnects which results in insufficient voltage over The simulation results for PUF UF as a function of yield
crosspoint devices. As a result, some of the devices will have before and after applying the balancing heuristic are plotted
greater current reduction resulting in an undesired bias which in Fig. 10(a) and (b), respectively. As shown in Fig. 10(a),
reduces PUF reliability. UF deviates from the ideal 50% when yield decreases. The
This article studies the effect of another hardware imper- UF somewhat improves for lower yield scenarios, likely due
fection on PUF, namely nonideal yield. When the yield is to the same assumed conductance values of stuck-on devices.
not 100%, some faulty devices exist that are stuck- at either Furthermore, the simulation results for PUF UF as a func-
ON - or OFF -states. When a device is stuck-at ON / OFF , its tion of crossbar size before and after applying the balanc-
current is much higher/smaller than that of other devices. ing heuristic are plotted in Fig. 10(c) and (d), respectively.
Authorized licensed use limited to: University of Massachusetts Amherst. Downloaded on August 24,2022 at 14:59:20 UTC from IEEE Xplore. Restrictions apply.
1822 IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 69, NO. 4, APRIL 2022
As shown in Fig. 10(c), the UF will be more resistant to [3] H. Nili et al., “Hardware-intrinsic security primitives enabled by ana-
hardware imperfections caused by the inferior yield when logue state and nonlinear conductance variations in integrated memris-
tors,” Nature Electron., vol. 1, no. 3, pp. 197–202, Mar. 2018.
size increases. In fact, when the PUF size increases, the [4] Y. Yoshimoto, Y. Katoh, S. Ogasahara, Z. Wei, and K. Kouno,
CRP space increases exponentially. Therefore, it will be less “A ReRAM-based physically unclonable function with bit error rate
probable for the stuck-at faulty device to be selected. Addi- <0.5% after 10 years at 125 ◦ C for 40 nm embedded application,”
in Proc. IEEE Symp. VLSI Technol., Honolulu, HI, USA, Jun. 2016,
tionally, as shown in Fig. 10(b) and (d), the balancing heuristic pp. 1–2.
improves PUF UF by setting to lower conductances the devices [5] M. R. Mahmoodi, H. Nili, and D. B. Strukov, “RX-PUF: Low power,
on the same lines with stuck-at ON devices. dense, reliable, and resilient physically unclonable functions based on
analog passive RRAM crossbar arrays,” in Proc. VLSI Symp., Honolulu,
HI, USA, Jun. 2018, Art. no. 176.
VII. D ISCUSSION AND C ONCLUSION [6] J. Kim et al., “A physical unclonable function with redox-based
nanoionic resistive memory,” IEEE Trans. Inf. Forensics Security,
The simulation results confirm that the robustness to vol. 13, no. 2, pp. 437–448, Feb. 2018.
machine learning attacks is better for analog nonlinear mem- [7] R. Liu, H. Wu, Y. Pang, H. Qian, and S. Yu, “Experimental character-
ization of physical unclonable function based on 1 kb resistive random
ristive PUFs as compared to their linear, digital coun- access memory arrays,” IEEE Electron Device Lett., vol. 36, no. 12,
terpart. The robustness is improved by decreasing device pp. 1380–1383, Dec. 2015.
conductance variations. These results are relevant for both [8] R. Liu, H. Wu, Y. Pang, H. Qian, and S. Yu, “A highly reliable and
tamper-resistant RRAM PUF: Design and experimental validation,” in
fixed-conductance and adjustable-conductance (i.e., reconfig- Proc. HOST, McLean, VA, USA, May 2016, pp. 13–18.
urable) PUFs [3], [11]. For a larger device variations in the [9] Y. Pang et al., “A reconfigurable RRAM physically unclonable function
crossbar circuit, we considered a scenario when the devices utilizing post-process randomness source with <6×10−6 native bit error
rate,” in IEEE Int. Solid-State Circuits Conf. (ISSCC) Dig. Tech. Papers,
can be selectively tuned with higher precision and proposed a San Francisco, CA, USA, Feb. 2019, pp. 402–404.
conductance balancing technique to boost robustness. A prac- [10] J. Yang et al., “A physically unclonable function with BER <0.35% for
tically valuable extension of such technique is to update the secure chip authentication using write speed variation of RRAM,” in
Proc. ESSDERC, Dresden, Germany, Sep. 2018, pp. 54–57.
optimal distribution of conductances based on the measured [11] M. R. Mahmoodi, H. Nili, Z. Fahimi, S. Larimian, H. Kim,
states of the already tuned devices and those of stuck-at-fault and D. Strukov, “Ultra-low power physical unclonable function with
devices. nonlinear fixed-resistance crossbar circuits,” in IEDM Tech. Dig.,
San Francisco, CA, USA, Dec. 2019, pp. 30.1.1–30.1.4.
More balanced device conductances and reduced σ naturally [12] A. Maiti, V. Gunreddy, and P. Schaumont, “A systematic method to
lead to a narrower distribution of differential currents IL − IR , evaluate and compare the performance of physical unclonable functions,”
which is the main challenge of hardware implementation. in Embedded Systems Design With FPGAs. New York, NY, USA:
Springer, 2013, pp. 245–267.
It is noted that the impact of reduced σ and conductance [13] X. Xi, A. Aysu, and M. Orshansky, “Fresh re-keying with strong
balancing is not just in the scaling of such distribution but PUFs: A new approach to side-channel security,” in Proc. IEEE Int.
also in making the responses less correlated, as confirmed Symp. Hardw. Oriented Secur. Trust (HOST), Washington, DC, USA,
Apr. 2018, pp. 1–4.
by modeling PUF responses with MLP networks. The tighter [14] G. T. Becker et al., “Active and passive side-channel attacks on delay
margins in reading differential currents, however, may degrade based PUF designs,” IACR Cryptol. ePrint Arch., vol. 2014, p. 287,
reliability. This issue is currently neglected in our study, Apr. 2014.
[15] X. Xu and W. Burleson, “Hybrid side-channel/machine-learning attacks
in part due to the assumption of ideal peripheral circuitry. The on PUFs: A new threat?” in Proc. Design, Autom. Test Eur. Conf. Exhib.
investigation of the reliability/robustness tradeoff is the next (DATE), Dresden, Germany, 2014, pp. 1–6.
important immediate goal. [16] U. Ruhrmair et al., “Power and timing side channels for PUFs and
their efficient exploitation,” IACR Cryptol. ePrint Arch., Tech. Rep. 851,
2013. [Online]. Available: https://eprint.iacr.org/2013/851
ACKNOWLEDGMENT [17] A. N. Mahmoud, U. Rührmair, M. Majzoobi, and F. Koushanfar,
“Combined modeling and side channel attacks on strong PUFs,” IACR
The authors would like to thank Mahdi Nazemi for useful Cryptol. ePrint Arch., vol. 2013, p. 632, Oct. 2013.
discussions and technical support. [18] Y. Gao, D. C. Ranasinghe, S. F. Al-Sarawi, O. Kavehei, and D. Abbott,
“Emerging physical unclonable functions with nanotechnology,” IEEE
Access, vol. 4, pp. 61–80, 2016.
R EFERENCES [19] Y. Pang, B. Gao, B. Lin, H. Qian, and H. Wu, “Memristors for hardware
security applications,” Adv. Electron. Mater., vol. 5, no. 9, Sep. 2019,
[1] M. R. Mahmoodi, D. B. Strukov, and O. Kavehei, “Experimental Art. no. 1800872.
demonstrations of security primitives with nonvolatile memories,” IEEE [20] P.-Y. Chen, R. Fang, R. Liu, C. Chakrabarti, Y. Cao, and S. Yu,
Trans. Electron Devices, vol. 66, no. 12, pp. 5050–5059, Dec. 2019. “Exploiting resistive cross-point array for compact design of physical
[2] Y. Gao, S. F. Al-Sarawi, and D. Abbott, “Physical unclonable functions,” unclonable function,” in Proc. IEEE Int. Symp. Hardw. Oriented Secur.
Nature Electron., vol. 3, no. 2, pp. 81–91, 2020. Trust (HOST), Washington, DC, USA, May 2015, pp. 26–31.
Authorized licensed use limited to: University of Massachusetts Amherst. Downloaded on August 24,2022 at 14:59:20 UTC from IEEE Xplore. Restrictions apply.