Toward remote and secure authentication: Disambiguation of magnetic microwire signatures using neural networks

Vol.:(0123456789)

MRS COMMUNICATIONS · VOLUME XX · ISSUE xx · www.mrs.org/mrc

MRS Communications

https://doi.org/10.1557/s43579-022-00302-5

Research Letter

Toward remote andsecure authentication: Disambiguation ofmagnetic

microwire signatures using neural networks

AksharVarma , Khoury College ofComputer Sciences, Northeastern University, Boston, MA, USA

XiaoyuZhang, Department ofMechanical andIndustrial Engineering, Northeastern University, Boston, MA, USA

BrianLejeune, Department ofChemical Engineering, Northeastern University, Boston, MA, USA

LauraCebadaAlmagro, Department ofMaterials Physics, Complutense University ofMadrid, Madrid, Spain

RafaelP.delReal, Instituto de Ciencia de Materiales de Madrid, Cantoblanco, Madrid, Spain; Instituto de Magnetismo Aplicado, Universidad Complutense de

Madrid-ADIF, Madrid, Spain

PilarMarin, Department ofMaterials Physics, Complutense University ofMadrid, Madrid, Spain; Instituto de Magnetismo Aplicado, Universidad Complutense de

Madrid-ADIF, Madrid, Spain

OgheneyunumeFitchorova, The George J. Kostas Research Institute forHomeland Security, Northeastern University, Burlington, MA, USA; KRI atNortheastern

University, LLC, Burlington, MA, USA; Department ofElectrical andComputer Engineering, Northeastern University, Boston, MA, USA

LauraH.Lewis, Department ofMechanical andIndustrial Engineering, Northeastern University, Boston, MA, USA; Department ofChemical Engineering, Northeastern

University, Boston, MA, USA; The George J. Kostas Research Institute forHomeland Security, Northeastern University, Burlington, MA, USA

RaviSundaram, Khoury College ofComputer Sciences, Northeastern University, Boston, MA, USA

Address all correspondence to Akshar Varma at [email protected]

(Received 5 July 2022; accepted 21 October 2022)

Abstract

Secure and high-throughput authentication systems require materials with uniquely identiable responses that can be remotely detected and rapidly

disambiguated. To this end, complex electromagnetic responses from arrangements of amorphous ferromagnetic microwires were analyzed using

machine learning. These novel materials deliver maximal spectral dispersion when the frequency of incident electromagnetic radiation matches the

microwire resonance. Utilizing data obtained from 225 unique microwire arrangements, a neural network reproduced the response distribution of

unseen data to a condence level of 90%, with a mean square error less than 0.01. This favorable performance afrms the potential of magnetic

microwires for use in tags for secure article surveillance systems.

Introduction

Secure, high-throughput, and contactless tracking of assets is

an enormous challenge for supply chain management as well

as for monitoring and controlling illicit transactions involv-

ing counterfeit currency and documents, pirated goods, and

substandard/falsied pharmaceuticals.

[1,2]

It is clear that sus-

tained advances in materials, devices, and data handling are

needed to address this complex and escalating issue. While

radio-frequency identication (RFID) technologies, including

chipless RFID,

[3]

can satisfy some of these requirements, their

expanded deployment remains constrained by cost, physical

size, and diculties associated with storing sucient data to

uniquely identify the signatures of extremely large numbers of

individual objects. At their essence, these identication systems

consist of an electromagnetic (EM)-active tag, a transmitter

to deliver interrogating incident EM radiation to the tag, and

a receiver to detect and analyze the subsequently emitted EM

signal. Although increased complexity in the received EM

signals, or spectra, provides a larger portfolio of potentially

unique tag signatures, at the same time it greatly increases the

challenge of disambiguating these signals and validating the

objects of interest.

To address this imperative, a system encompassing creation,

interrogation, and disambiguation of complex electromagnetic

signatures is described here. Signals emitted from EM-active

“tags” comprising specied arrangements of unique micron-

scaled magnetic objects—amorphous magnetic microwires—

were analyzed using machine learning in the form of neural

networks.

[4]

Results reported here were obtained from a surpris-

ingly small number (225) of measurements, and these results

conrmed parameterized learning of the response function to

successfully reproduce training and testing set data to a com-

mendable condence level in excess of 90%. In this manner, a

proof-of-concept demonstration has been achieved, allowing

contemplation of strategies to rene the parameter space and

physical conguration of the magnetic tags for improved per-

formance and applicability.

The uniqueness of this current work is derived from the

application and integration of three typically separate knowl-

edge domains—advanced magnetic materials,

[5]

machine learn-

ing,

[6]

and information theory

[7]

—to address a complex systems

challenge. Below we briey elaborate on the novelty and sig-

nicance of combining these three disciplinary arenas. As an

innovative composite material, glass-coated magnetic microw-

ires have the ability to modulate reection or transmission of

microwave (GHz) EM radiation incident on their surface, with

maximum dispersion achieved when the incident frequency is

matched to the microwire antenna resonance frequency. The

MRS COMMUNICATIONS

· VOLUME XX · ISSUE xx · www.mrs.org/mrc

emitted microwire signal is very sensitive to details of chemical

composition as well as to environmental parameters (including

ambient magnetic eld, strain, and temperature).

[8,9]

Further,

the response of a multiwire ensemble is inuenced by the num-

ber, length, and proximity of the constituent wires.

[10,11]

Overall

this abundance of controllable parameters provides an immense

and highly versatile palette of variables and conditions to real-

ize unique tracking “tags” comprised of magnetic microwires.

The high-fidelity reproduction by a neural network of the

electromagnetic response of these magnetic microwire arrays

allows us to envision simulating a variety of materials composi-

tions in conjunction with physical congurations, to eciently

explore and survey a much larger tag space.

[12,13]

It is notable

that our neural networks achieve good performance with train-

ing dataset sizes on the order of a few hundreds, while machine

learning and especially neural networks typically require data-

set sizes on the order of 10,000 to a few million; this aspect is

noted within the materials science community as well.

[14]

With

the tag modeled as a neural network, we are now in a position

to innovate on the neural network architecture front to create

ecient encoder and decoder for tags.

[15,16]

In follow-on pre-

liminary work,

[17]

we leverage the proof-of-concept described

in this paper to create a general technique for creating a practi-

cal tagging system from any scanning technology employing a

novel autoencoder-based neural network architecture.

Materials andmethods

Magnetic microwires andtheir response

Amorphous glass-coated ferromagnetic microwires derive their

very high sensitivity to applied magnetic elds

[18–20]

from inter-

play between their atomic and magnetic structures.

[21]

In addi-

tion to ultrasoft magnetic behavior (coercivity H

∼

0.2 Oe),

these materials can exhibit a giant magnetoimpedance (GMI)

eect or a ferromagnetic resonance (FMR) eect;

[22–24]

in the

absence of applied electric current, this phenomenon is referred

to as the “antenna eect.”

[22–24]

Here, investigation is focused

on the antenna eect response of amorphous magnetic microw-

ires comprising of an amorphous metallic ferromagnetic core

(typical radius 10–20 microns) covered by uniform borosilicate

glass (i.e., pyrex) coating of approximately 20–30 microns in

thickness. The microwires of this study were procured from

Micror Tehnologii Industriale Ltd. (Moldova) with a consist-

ent composition of (Co

0.94

0.06

)

and an outer wire

diameter of 75–100 microns. Data were collected from congu-

rations of parallel wire arrays (array specics provided in the

following section) that were mounted on a piece of dielectric

plastic lm and secured with clear adhesive tape.

The initial neural network computational analyses were

performed on 225 separate and unique measurements (aka

congurational response pairs) of the antenna eect response

collected from the glass-coated amorphous magnetic microw-

ires arranged in a variety of configurations as tags. The

microwire arrays were assessed in the frequency domain in the

range 1–4 GHz using two double-ridged guide horn antennas

(ETS-LINDGREN model 3115) as an emitter and a receiver;

these were spaced 1.2 m apart to ensure a far-eld congura-

tion. The antennas were connected to a programmable network

analyzer (Agilent E8362B PNA Series Network Analyzer). The

microwire arrays were oriented perpendicular to the midline

connecting the two antennas, perpendicular to the direction of

the incident EM waves. After open-air calibration, the scatter-

ing parameters

, which quantify in decibels (dB) the ratio of

the emitting antenna power (

) to the receiving antenna power

(

), were determined according to Equation 1:

The resultant EM scattering information was collected in the

frequency domain where the measured

scattering coecient

presents a minimum.

Neural network architecture andtraining

Using the data collected as described above, a neural-network-

based machine learning model

[4,25]

was developed and opti-

mized to predict the

response generated by a given congu-

ration of magnetic microwires. The microwire congurations

used to generate signals were initially dened using three fea-

tures: (a) the length of the microwires (either 3, 4, or 8 cm), (b)

the number of the microwires (between 1 and 16), and (c) the

separation between microwires (ranging from 0.4 to 20 cm).

The response of a particular conguration was represented by a

200-dimensional vector of dB values denoting the

response,

with each dB value corresponding to xed, linearly separated

frequencies in the 1–4 GHz range.

These three features describing the microwire congurations

were input into the neural network model with a single hidden

layer of 1000 articial neurons with rectied linear unit (ReLU)

activations.

[26,27]

In the eld of neural networks, the term “neu-

rons” refer to an operation that performs a weighted sum of all

the inputs, with the weights being parameters that are optimized

using training data. Batch normalization

[28]

was applied to the

hidden layer before the output layer produced an output of a

200-dimensional vector designed to match the recorded

response. The model was trained on Mean Square Error (MSE)

loss using the popular ADAM optimizer,

[29]

with a learning

rate hyperparameter value of 1e−3 and a L2 regularization

[30]

hyperparameter value of 1e−6 for up to 2000 epochs. The MSE

represents a statistical assessment of the validity of the model

and is the average of the square of the dierences between the

target dB values and predicted dB values, which is minimized

during the training phase. From the total 225 (conguration,

response) pairs that were initially studied, a randomly chosen

selection of 202 measurements (90% of the total number) was

used for training the neural network model, while the remaining

23 measurements (10% of the total number) were employed as

“unseen” data to evaluate the model’s performance. This 90-10

split was performed 10 times to quantify the error associated

with the random splitting of the training and testing data.

(1)

= 20 · log

Research Letter

MRS COMMUNICATIONS · VOLUME XX · ISSUE xx · www.mrs.org/mrc

Results anddiscussion

Initial outcomes

The neural network favorably reproduced the response distribu-

tion of unseen data to a condence level of 90%, with a quite

favorable mean square error less than 0.01. The model out-

comes are provided in the rst two columns of Table I that con-

tain the MSE values resulting from both the training and testing

phases. Corresponding plots of the actual and the predicted

responses obtained from a selection of the unseen microwire

congurations are provided in Fig. 1. Good agreement between

the actual data and the predicted data is visually evident, and is

anticipated to improve further as more congurational response

pairs are analyzed by this neural network model. The high qual-

ity of the predictions, as obtained from only 202 separate spec-

tral response measurements, is very noteworthy as conventional

machine learning wisdom indicates that thousands of images

are typically necessary for achievement of good predictions.

[31]

Calibrating theneural network model

fornew environments

To evaluate the performance of the model in evaluating

microwire array data collected in new environments arising

from altered measurement conditions, eects of calibrating,

or ne-tuning, the neural network was investigated using a

small subset of the additional measurements (referred to as

the calibration set). Of particular note are the dierences in

how well shielded the measurement apparatus were to the

ambient magnetic elds and temperatures, allowing us to

understand our model’s performance in more real-life envi-

ronments as opposed to a shielded lab environment. In the

absence of this calibration step, the performance of the origi-

nal neural network applied to the additional data obtained in

a new environment was noted to degrade; that is, model pre-

dictions of the microwire array response measured in a new

environment resulted in greatly increased MSE values (see

Table I, Column 3). To perform this ne-tuning, the original

neural network model was trained using this calibration set

Figure1. Actual responses (blue traces) and predicted responses (orange traces) for the scattering coefcient data (

) originating from

various unseen magnetic microwire tag congurations.

Table I. Resultant mean square error (MSE) values and correspond-

ing error in magnetic microwire transmission coecient responses

(

The neural network model was developed from 202 randomly

chosen training measurements and was tested on 23 unseen

measurements. The nal two columns in this Table denote the

performance of the neural network in the new environment

(a dierent measurement condition) both without and with

calibration/ne-tuning.

Original training

MSE value

Original testing

MSE value

Testing MSE

value of original

model applied

to data obtained

in a new envi-

ronment

Testing MSE

value of ne-

tuned model

applied to data in

a new environ-

ment

0.002 ± 0.0005 0.0075 ± 0.0007 0.02 ± 0.009 0.008 ± 0.0006

MRS COMMUNICATIONS

· VOLUME XX · ISSUE xx · www.mrs.org/mrc

for up to 2000 epochs, as was conducted earlier, with the

other training hyper-parameters kept xed as described in the

previous section. It was found that calibration sets consist-

ing of as few as 30 additional measurements (compared to

the 202 measurements required for the initial training activ-

ity) were sucient to enable the model to detect nuances

of responses to the new environment and improve the MSE

values, restoring them back to their original low levels (see

Table I, Column 4). Note that quantication of error in the

MSE values, arising from random selection of the original

training data or the calibration set, is obtained from 10 rep-

etitions of the ne-tuning step, each conducted with a dif-

ferent random calibration set. This allows us to average out

uctuations in the results due to the randomness in the neural

network training process, and we report the observed uctua-

tions as a condence interval around the average MSE values.

Analyzing thesensitivity ofmachine

learning models

Once the models were able to accurately predict responses

obtained from various magnetic microwire tag congurations

and adapt to changing environments, the sensitivity of the

machine learning models to dierences in these congura-

tions was investigated. That is, it was desired to determine

the smallest change in response that the machine learning

model could correctly accommodate. This aspect was inves-

tigated by acquiring and analyzing new datasets that probed

three additional microwire conguration features. The rst

additional feature was the angle

between the wire orienta-

tion and line connecting the antennas; this feature tracked

in-plane changes in the direction of the emitted EM wave.

The remaining two additional features, expressed as two-

dimensional x and y coordinates, tracked the in-plane position

of the microwire relative to the midline position between the

emitter and receiver antennas. The x-shift described the dis-

tance that the wires were displaced along the direct transmit-

ter/receiver line (closer to/farther away from the antennas).

The y-shift was perpendicular to the x-shift.

Preliminary observations indicate that the spectral predic-

tions returned by the machine learning models are poor when

the angle

is changed from its original value of zero degrees.

We hypothesize that these poor results are due to the response

changing as the wires move away from being perpendicular

to the direction of the electromagnetic wave propagation. The

machine learning models, in turn, could not compensate for

this change in the nature of the response. The spectral predic-

tions of the machine learning models are better for the x-shift;

however, even with respect to that feature, the MSE values

were worse than those reported in previous sections. While

the reasons underlying these results are not clear at the current

time, it is believed that they may have their origins in details of

the detected response in near-eld versus far-eld conditions.

In contrast, spectra produced from the y-shift is easy for the

model to predict, with MSE values in the same range as the

results depicted in Table I. In fact, signicant dierences in

responses produced by microwire arrays that have been shifted

only in the y-direction are quite distinguishable.

Conclusions

Successful prediction (

< 0.01

mean squared error (MSE),

within 90% condence level) of “unseen” high-frequency

electromagnetic responses measured from 2D arrays of amor-

phous ferromagnetic microwires was achieved using neural

network-based machine learning. These results, obtained

from a surprisingly small number (225) measurements, were

extended using an additional 30 measurements to ne-tune

the model for improved robustness to varied environments.

This work combines magnetic materials science, speci-

cally the electromagnetic response of amorphous magnetic

microwires, with machine learning techniques for training

neural networks to faithfully reproduce the microwire GHz

antenna responses with high delity. We demonstrate that

with the carefully chosen neural network architectures and

clean data, it is possible to achieve good performance using

very few measurements compared to what is considered the

norm in neural network literature.

[31]

Interpreting the results

of the machine learning model to improve the physical under-

standing of the properties of magnetic microwire signatures

is part of ongoing and future work. Considering the overall

abundance of controllable parameters enabled by the mag-

netic microwires, we have an immense and highly versatile

palette of variables and conditions to realize unique track-

ing “tags.” Combined with the high-delity reproduction by

a neural network of the electromagnetic response of these

magnetic microwire arrays allows us to envision simulat-

ing a variety of materials compositions alongside varying

physical congurations of these materials, to eciently cap-

ture and understand large scale tag spaces.

[12,13]

Our results

allow us to model tags using a neural network, which allow

innovative neural network architectures to create ecient

encoder and decoder for tags to and from their electromag-

netic responses.

[15,16]

In follow-on preliminary work,

[17]

by employing a novel autoencoder architecture along with

concepts from information theory, we leverage the proof-of-

concept described in this paper to create a general algorithm

for creating a practical tagging system from any scanning

technology.

These proof-of-concept results are but the necessary rst

step in a novel approach toward the eventual goal of hid-

den machine-readable authentication of objects (equipment,

medical and pharmaceutical products, materials, documents,

currency, etc.). We note that the security of supply chains in

the USA is a market estimated to be over a billion dollars in

size in 2021 and growing at a CAGR greater than

. Our

approach and architecture are very general. With any scan-

ning technology modeled as a deep network, we envision a

cyber-physical system—the cyber half constitutes the brains

exploring and organizing the tag space, while the physical

Research Letter

MRS COMMUNICATIONS · VOLUME XX · ISSUE xx · www.mrs.org/mrc

robotic half constitutes the brawn, creating and validating

actual physical tags, and the two work in close coordination

to create an ecient and practically usable tag system.

Funding

This study was funded by Northeastern University, NSF D-ISN

2039945, Fulbright España, I-Link A20074 (CSIC), Span-

ish Ministry of Science and Innovation RTI2018-095856-B-

C21 and Comunidad de Madrid NANOMAGCOST S2018/

NMT-4321.

Data availability

The authors will not make data used for training the neural net-

work models available since they are part of an IP disclosure.

Code availability

The code used for analysis will be made available on request.

Declarations

Conﬂict of interest

The authors have no conicts of interest.

References

1. T.K. Mackey, G. Nayyar, A review of existing and emerging digital tech-

nologies to combat the global trade in fake medicines. Expert Opin. Drug

Saf. 16(5), 587–602 (2017)

2. M. Attaran, Digital technology enablers and their implications for supply

chain management. Supply Chain Forum 21, 158–172 (2020)

3. A. Subrahmannian, S.K. Behera, Chipless rd: a unique technology for

mankind. IEEE J. Radio Freq. Identif. (2022). https:// doi. org/ 10. 1109/

JRFID. 2022. 31469 02

4. I.J. Goodfellow, Y. Bengio, A. Courville, Deep learning (MIT Press, Cambridge,

2016)

5. M. Vazquez, Magnetic nano- and microwires design, synthesis, properties

and applications (Elsevier, Amsterdam, 2015). https:// www. elsev ier. com/

books/ magne tic- nano- and- micro wires/ vazqu ez/ 978-0- 08- 100164-6

6. M. Mohri, A. Rostamizadeh, A. Talwalkar, Foundations of machine learning

(The MIT Press, Cambridge, 2012)

7. T.M. Cover, J.A. Thomas, Elements of information theory (Wiley series in

telecommunications and signal processing) (Wiley-Interscience, Hoboken,

2006)

8. V.V. Popov, V.N. Berzhansky, H.V. Gomonay, F.X. Qin, Stress-induced mag-

netic hysteresis in amorphous microwires probed by microwave giant

magnetoimpedance measurements. J. Appl. Phys. 113(117), 17A326

(2013)

9. D. Makhnovskiy, A. Zhukov, V. Zhukova, J. Gonzalez, Tunable and self-sensing

microwave composite materials incorporating ferromagnetic microwires.

Adv. Sci. Technol. 54, 201–210 (2008)

10. D. Archilla, A. Hernando, E. Navarro, Boosting the tunable microwave

scattering signature of sensing array platforms consisting of amorphous

ferromagnetic Fe

2.25

72.75

microwires and its amplication by

intercalating Cu microwires. Nanomaterials 11, 1–16 (2021)

11. A. Uddin, D. Estevez, F.X. Qin, From functional units to material design: a

review on recent advancement of programmable microwire metacom-

posites. Composites A 153, 106734 (2022)

12. G.B. Goh, N.O. Hodas, A. Vishnu, Deep learning for computational chem-

istry. J. Comput. Chem. 38(16), 1291–1307 (2017)

13. Y. Kim, Y. Kim, C. Yang, K. Park, G.X. Gu, S. Ryu, Deep learning framework

for material design space exploration using active transfer learning and

data augmentation. NPJ Comput. Math. 7, 140 (2021)

14. M.L. Pasini, P. Zhang, S.T. Reeve, J.Y. Choi, Multi-task graph neural net-

works for simultaneous prediction of global and atomic properties in

ferromagnetic systems. Mach. Learn.: Sci. Technol. 3(2), 025007 (2022)

15. H. Kim, Y. Jiang, R. Rana, S. Kannan, S. Oh, P. Viswanath, Communication

algorithms via deep learning (2018). arXiv preprint arXiv: 1805. 09317

16. H. Kim, Y. Jiang, S. Kannan, S. Oh, P. Viswanath, Deepcode: feedback

codes via deep learning. IEEE J. Sel. Areas Inf. Theory 1(1), 194–206

(2020)

17. R. Sundaram, A. Varma, Dispersive Autoassociative Neural Networks with

Inversion (DANNI) for generating constrained codes. Northeastern techni-

cal report (2022)

18. H. Wiesner, J. Schneider, Magnetic properties of amorphous Fe-P alloys

containing Ga, Ge, and As. Phys. Status Solidi (a) 26(1), 267 (1974)

19. J. Schneider, H. Wiesner, R. Gemperle, Annealing effects on the magnetic

properties of rapidly quenched transition metal alloys. Phys. Status Solidi

(a) 36(1), 59–64 (1976)

20. R. Gemperle, L. Kraus, J. Schneider, Magnetization reversal in amor-

phous (Fe1- xNix) 80P10B10 microwires. Czechoslov. J. Phys. B 28(10),

1138–1145 (1978)

21. M. Vazquez, Advanced magnetic microwires. Handb. Magn. Adv. Magn.

Mater. (2007). https:// doi. org/ 10. 1002/ 97804 70022 184. hmm418

22. P. Marín, M. Marcos, A. Hernando, High magnetomechanical coupling on

magnetic microwire for sensors with biological applications. Appl. Phys.

Lett. 96(26), 262512 (2010)

23. C. Herrero-Gómez, P. Marín, A. Hernando, Bias free magnetomechanical

coupling on magnetic microwires for sensing applications. Appl. Phys.

Lett. 103(14), 142414 (2013)

24. A. Hernando, V. Lopez-Dominguez, E. Ricciardi, K. Osiak, P. Marin, Tuned

scattering of electromagnetic waves by a nite length ferromagnetic

microwire. IEEE Trans. Antennas Propag. 64(3), 1112–1115 (2015)

25. C.M. Bishop, Neural networks for pattern recognition (Oxford University

Press, Oxford, 1995)

26. K. Fukushima, Cognitron: a self-organizing multilayered neural network.

Biol. Cybern. 20(3), 121–136 (1975)

27. V. Nair, G.E. Hinton, Rectied linear units improve restricted boltzmann

machines, in Proceedings of the 27th international conferenceon machine

learning. (Omnipress, Haifa, 2010), pp.807–814

28. S. Ioffe, C. Szegedy, Batch normalization: accelerating deep network

training by reducing internal covariate shift, in International conference

on machine learning. (PMLR, 2015), pp. 448–456

29. D.P. Kingma, J. Ba, Adam: a method for stochastic optimization (2014).

arXiv preprint arXiv: 1412. 6980

30. M. Arashi, A.M.E. Saleh, B.G. Kibria, Theory of ridge regression estimation

with applications (Wiley, New York, 2019)

31. Wikipedia, List of datasets for machine-learning research. Wikipedia

(2022). https:// en. wikip edia. org/ wiki/ List_ of_ datas ets_ for_ machi ne-

learn ing_ resea rch

Publisher’s Note Springer Nature remains neutral with regard to

jurisdictional claims in published maps and institutional aliations.