PyCRA: Physical Challenge-Response Authentication For Active Sensors Under Spoofing Attacks

Transcription

1 PyCRA: Physical Challenge-Response Auhenicaion For Acive s Under Spoofing Aacks Yasser Shoukry, Paul Marin, Yair Yona, Suhas Diggavi, and Mani Srivasava Elecrical Engineering Deparmen, Universiy of California a Los Angeles, USA {yshoukry, pdmarin, yairyo99, suhasdiggavi,mbs}@ucla.edu ABSTRACT Embedded sensing sysems are pervasively used in life- and securiycriical sysems such as hose found in airplanes, auomobiles, and healhcare. Tradiional securiy mechanisms for hese sensors focus on daa encrypion and oher pos-processing echniques, bu he sensors hemselves ofen remain vulnerable o aacks in he physical/analog domain. If an adversary manipulaes a physical/analog signal prior o digiizaion, no amoun of digial securiy mechanisms afer he fac can help. Forunaely, naure imposes fundamenal consrains on how hese analog signals can behave. This work presens PyCRA, a physical challenge-response auhenicaion scheme designed o proec acive sensing sysems agains physical aacks occurring in he analog domain. PyCRA provides secure acive sensing by coninually challenging he surrounding environmen via random bu deliberae physical probes. By analyzing he responses o hese probes, he sysem is able o ensure ha he underlying physics involved are no violaed, providing an auhenicaion mechanism ha no only deecs malicious aacks bu provides resilience agains hem. We demonsrae he effeciveness of PyCRA in deecing and miigaing aacks hrough several case sudies using wo sensing sysems: (1) magneic sensors like hose found on gear and wheel speed sensors in roboics and auomoive, and (2) commercial Radio Frequency Idenificaion (RFID) ags used in many securiy-criical applicaions. In doing so, we evaluae boh he robusness and he limiaions of he PyCRA securiy scheme, concluding by oulining pracical consideraions as well as furher applicaions for he proposed auhenicaion mechanism. Caegories and Subjec Descripors C.2. [COMPUTER-COMMUNICATION NETWORKS]: General: Securiy and proecion General Terms Securiy Permission o make digial or hard copies of all or par of his work for personal or classroom use is graned wihou fee provided ha copies are no made or disribued for profi or commercial advanage and ha copies bear his noice and he full ciaion on he firs page. Copyrighs for componens of his work owned by ohers han ACM mus be honored. Absracing wih credi is permied. To copy oherwise, or republish, o pos on servers or o redisribue o liss, requires prior specific permission and/or a fee. Reques permissions from Permissions@acm.org. CCS 1, Ocober 12 16, 21, Denver, Colorado, USA. c 21 ACM. ISBN /1/1...$1.. DOI: hp://dx.doi.org/1.114/ Keywords Embedded Securiy; Acive sensors; Challenge-response auhenicaion; Spoofing aacks; Physical aacks 1. INTRODUCTION Recen decades have winessed a proliferaion in embedded sensors for observing a variey of physical phenomena. Increased use of hese sensors in securiy- and life-criical applicaions has been accompanied by a corresponding increase in aacks argeing sensing sofware, hardware, and even physical, analog signals hemselves. While considerable research has explored sensor securiy from a sysem-level perspecive nework redundancy, sensor fusion, and he like sensors hemselves remain largely vulnerable o aacks argeing analog signals prior o digiizaion. This vulnerabiliy can lead o caasrophic failures when a malicious hird pary aemps o spoof he sensor [19, 14, 3, 33]. Several sysem-level sensor securiy schemes have been proposed in he conex of power grids. For example, Dorfler e al. have explored disribued cyber-physical aack deecion in he conex of power neworks [6, 3]. Similar ideas for providing sysem-level securiy in smar grids can be found in [16, 18, 4, 22, 3]. Securiy schemes in his vein include, among ohers, sae-space and conrolheoreic approaches for deecing anomalous sysem behavior [7, 3]. One idea common o hese effors is ha an inheren securiy mechanism and robusness can be found in he physics governing he dynamics of he sysem as a whole. For example, a mismach beween he rae of change in a vehicle s locaion as repored by GPS and by he odomeer sensor may indicae ha one of hese wo sensors is eiher fauly or under aack. A complemenary securiy mechanism can be found in he underlying physics governing he sensor iself. If a sensor observes an analog signal ha appears o violae he physics governing he sensing dynamics, he signal iself may be under aack, necessiaing securiy mechanisms a he analog signal level. To reduce sensor-level vulnerabiliies, engineers ofen place sensors in secure or remoe physical locaions o preclude direc physical conac wih he sensing hardware. Addiionally, he phenomenon being sensed is ofen difficul o access, wheher prohibiively far away or surrounded by proecive maerial. In such scenarios, adversaries have access only o he analog signal prior o i reaching he sensor, and heir aack mus be carried ou wihou direc access o any hardware in he enire sensing pah, from source o sink. Even wih hese counermeasures in place, an adversary can sill aack sensors by manipulaing he physical signals before heir ransducion and subsequen digiizaion [19, 33]. Robus counermeasures for such aacks mus necessarily be carried ou a he physical level as well once hese signals have been sampled and digiized, no amoun of pos-processing can repair he compromised sensor daa. 14

2 Broadly speaking, sensors can be divided ino wo caegories: passive (hose ha sense pre-exising physical signals) and acive (hose ha perform some acion o evoke and measure a physical response from some measurable eniy). Examples of passive sensors include emperaure, humidiy, and ambien ligh, while acive sensors include ulrasound, laser scanners, and radar. Passive sensors are largely naïve lisening devices hey blindly relay informaion o higher levels of sofware wihou regard for he inegriy of ha informaion. Digial filering and oher pos-processing echniques can be used o remove noise from passive sensors, bu hey remain unable o comba aacks a he physical layer in any meaningful way. On he oher hand, acive sensors inroduce he possibiliy for more advanced securiy measures. PyCRA is, a is core, a mehod of ensuring he rusworhiness of informaion obained by acive sensors by comparing heir responses o a series of physical queries or challenges. The driving concep behind PyCRA is ha, by periodically simulaing he environmen wih a known signal and measuring he response, we can ensure ha he signal measured by he sensor is in accordance wih he underlying sensing physics. This periodic simulaion and subsequen behavioral analysis he physical challenge-response auhenicaion, creaing a secure acive sensing plaform is he main conribuion of his work. We demonsrae he effeciveness of PyCRA for hree exemplary cases: physical aack deecion for magneic encoders, physical aack resilience for magneic encoders, and passive eavesdropping deecion for RFID readers. Magneic encoders are used in a wide array of commercial and indusrial applicaions and are represenaive of a large class of inducive acive sensors. We demonsrae no only how acive spoofing aacks can be deeced for hese inducive sensors bu also how he effecs of hese aacks can be miigaed. Eavesdropping deecion on RFID readers serves o illusrae an exension of PyCRA o enable deecion of passive aacks. Our resuls from more han 9 experimens demonsrae ha PyCRA can accuraely deec aacks in a variey of seings. We believe ha he mehods demonsraed in his work can be applied o a broad array of acive sensors beyond hose sudied direcly in his work, including ulrasound, opical sensors, acive radar, and more. 1.1 Conribuions of PyCRA In summary, he conribuions described in his paper are hreefold: We presen a generalizable physical challenge-response auhenicaion scheme for acive sensing subsysems. We inroduce algorihms for deecing he presence of and providing resilience agains physical aacks when using physical challenge-response auhenicaion. We demonsrae he effeciveness of PyCRA, our implemenaion of physical challenge-response auhenicaion, agains several differen aack ypes and for over 9 experimens wih hree exemplary applicaions: (1) deecion of acive aacks on magneic encoders, (2) resilience agains acive aacks on magneic encoders, and (3) deecing passive eavesdropping aacks on RFID readers. The res of his paper is organized as follows. Secion 2 oulines he aacker model. Secion 3 describes he basic operaion of he PyCRA auhenicaion scheme for deecing acive aacks. Secion 4 oulines heoreical limiaions of aackers on physical signals. Secion, 6, and 7 are devoed o he resuls of hree case sudies: aack deecion for magneic encoders, aack deecion for passive eavesdropping on RFID readers, and aack resilience for magneic Sofware Layers Digial Back-End Acuaor Digiizaion (ADC) Acive Analog Fron-End (sensing & filering) Probe physical aack Response Figure 1: A ypical acive sensor archiecure. The acuaor generaes an analog signal (energy) which is refleced by he measured eniy back o he sensor. The received analog signal is capured and processed by he analog fron-end. The signal is hen convered o a digial forma which is processed once more (by he digial backend) before being sen o higher level sofware layers. encoders. Finally, we offer a discussion and concluding houghs in Secions 8.1, 8.2 and ATTACKER MODEL Before describing mechanisms by which we can deec and preven sensor aacks a he physical layer, we mus differeniae beween wo broad caegories of sensors namely passive and acive sensors and define wha we mean by a physical aack. 2.1 Passive vs. Acive s s can be broadly classified as eiher passive or acive based on he source of energy being sensed. Passive sensors measure ambien energy. For example, emperaure sensors like hose found in hermosas are considered passive, because hey measure hea energy in he ambien environmen. By conras, acive sensors probe some physical eniy wih self-generaed energy as shown in Figure 1. This energy is parially refleced back o he sensor where i is measured and used o infer properies abou some physical phenomenon. Examples of acive sensors include ulrasonic range finders (used in roboics), opical and magneic encoders (used in auomoive vehicles, indusrial plans, & chemical refineries), radar, and even radio-frequency idenificaion (RFID) sysems. In RFID, a reader is used o generae elecromagneic waves which are hen used by wireless ags o ransfer back heir unique idenifier o he reader. In his paper, we focus on providing securiy for acive sensors. In paricular, we leverage an acive sensor s abiliy o emi energy in order o 1) provide deecion of acive aackers rying o spoof he sensor, 2) miigae he effecs of acive spoofing aacks and 3) deec passive eavesdropping aacks aemping o lisen o he informaion received by he sensor. In he following subsecions, we define wha we mean by physical aacks on acive sensors and ouline he assumed properies and limiaions of a poenial adversary. 2.2 Defining Physical Aacks In his paper, a physical aack refers o a malicious aleraion of a physical, analog signal (e.g., magneic waves, acousic waves, visible waves) prior o ransducion and digiizaion by a sensor, as shown in Figure Adversarial Goals The adversary considered in his work has a number of goals relaed o misinforming and misleading sensors. These goals are summarized below. G1 Concealmen: An aacker does no wan he presence of his or her aack o be known. If a sensor aack can be easily deeced, prevenaive counermeasures like hardware redundancy and resilience a he sysem-level can ofen be used o miigae he damage done by he aack [7, 3]. Measured Eniy 1

3 Volage [vol] Physical/Analog signal 1 7 Logical represenaion Volage [vol] Physical/Analog signal 1 4 Logical represenaion Figure 2: Examples of physical delays seen in ypical sensing and acuaion hardware, including opical sensors (lef) and elecromagneic coupled (e.g., RFID) sensors (righ). In each case, he measured analog signal (blue solid) lags behind he ideal, logical signal (red dashed), causing delays. G2 Signal Injecion: An aacker will aemp o rick he sensor ino hinking ha a malicious, injeced signal is he rue physical signal. The primary goal of an aack is o replace he rue physical signal ha a sensor aims o sense wih a malicious signal. In oher words, an adversary will aemp o injec a signal ino he physical medium ha he sensor is measuring in order o jam or spoof he sensor. G3 Signal Masking: An aacker will aemp o preven he sensor from being able o deec he rue physical signal. If he sensor is sill capable of reliably discerning he correc signal from he malicious, injeced signal, hen he aack may no be successful. Thus, he adversary aims no only o injec a signal bu also o mask he rue signal, wheher by overpowering, modifying, or negaing (canceling) i. 2.4 Assumpions abou he Adversary The physical aacks agains sensors considered in his work operae under four main assumpions: A1 Non-invasiveness: Aacks are of a non-invasive naure ha is, he aacker is no allowed direc access o he sensor hardware. Addiionally, he adversary does no have access o he sensor firmware or sofware, wheher direcly or hrough wired or wireless neworking. In mos life- and safey-criical applicaions, engineers are careful o ensure ha sensors are no physically exposed and vulnerable o direc ampering. For example: s are ofen insalled inside he body of a physically secured infrasrucure (e.g., sensors inside he body of an auomoive sysem, moving UAV drones, ec.). For sensors which are physically accessible, exising echniques in he lieraure demonsrae ways o implemen amperproof packaging o proec sensors from direc, physical modificaions [31, 17, 1]. Numerous sensor sysems have mehods for deecing when wires connecing heir various sensors have been ampered wih. For example, auomoive sysems are equipped wih sensor failure deecion sysems which can deec wheher all sensor subsysems are correcly conneced and aler he driver if any of hem fails [8]. Because of his, any aack mus be carried ou from a disance, wihou direc access o any sensor hardware. In shor, an adversary is assumed o have access only o he physical/analog medium used by he sensor magneic waves, opics, acousics, ec. Addiionally, i is imporan o disinguish hese sensors from sensor nodes (which appear in he lieraure of sensor neworks); he aacks and counermeasures in his work arge sensors hemselves. s are simple subsysems designed o perform only one simple ask; sensing he physical world. Because of his, many sensors do no suppor remoe firmware updaes and do no ypically receive commands from a remoe operaor, making such aack vecors uncommon as many sensors do no have such capabiliies. A2 Trused Measured Eniy We assume ha he physical eniy o be measured by he sensor is rused and incapable of being compromised. Similar o he sensor hardware iself, he eniy ha he sensor aims o measure is ypically difficul o access or aler direcly while mainaining Goals G1 G3. For example, in RFID sysems he ag iself is ofen encased in amper-proof packaging [31, 17]; for ulrasonic ranging and acive radar, maliciously alering he measured eniy (ofen he enire surrounding environmen) is impracical in ime & effor and undoubedly violaes Goal G1; for airplane engine speed sensors, he engines canno easily be modified or replaced; for hear moniors, he hear canno (we hope) be modified [19], and so forh. A3 Physical Delays (τ aack ): Adversaries require physical hardware wih inheren physical delays. This delay, hough variable in duraion, is fundamenal o all physical acuaion and sensing hardware. These same analog/physical signals canno be manipulaed or even observed (i.e. sniffed) wihou physical hardware. Tha is, o amper wih magneic waves, an aacker needs hardware ha is able o generae magneic waves, opical signals need physical hardware ha generaes opical signals, and so on. Furhermore, his hardware has o obey fundamenal physics imposed by naure; he underlying physics dicae ha he response of any physical elemen is governed by a dynamical model (mahemaically modeled using differenial/difference equaions) [9, ch. 2], [2, chs. 8 9]. This dynamical model describes he oupu response for each physical elemen in response o heir inpus, e.g., he ime for a volage o drop from a cerain value o zero and so on. Alhough from a sysem poin of view, we ofen assume ha analog signals like hose in Figure 2 ake on logical values of and 1, he underlying physics is always differen from his sysem poin of view. For example, Figure 2 shows how hardware ha generaes clock waveforms and opical pulse signals behaves quie differenly from he desired, logical signals used o conrol hem. In general, no physical signal can arbirarily jump from one sae o anoher wihou suffering from delays imposed by physics [9, ch. 2]. Furhermore, hese physical delays are lower bounded by a nonzero, fundamenal limi. For example, he ime response of an elecromagneic sensor/acuaor is a muliple of physical consans like magneic permeabiliy [2, chs. 8 9] or permiiviy and elecric consans for capaciive sensors [9, ch. 4]. In general, he ime response of any sensor or acuaor can never be below cerain fundamenal hresholds conrolled by physical consans. We refer o his physical delay as τ aack for he remainder of his paper. A4 Compuaional Delays: PyCRA is designed and analyzed wih a focus on physical delays. We make no assumpion regarding he compuaional power of a poenial adversary. We assume ha an adversary has knowledge of he underlying securiy mechanism, aemping o conceal an aack by reacing o each physical challenge or probe from he PyCRA-secured acive sensor. In pracice, such an adversary would suffer from compuaional delays in addiion o he physical delays addressed above. These delays would make i even more difficul for an adversary o respond o hese challenges in a imely manner. PyCRA is designed 16

4 Acive Acuaor probe response Measured Eniy Malicious Acive Acuaor probe response Measured Eniy Malicious Acuaor Acive (a) (b) (c) Figure 3: An illusraion of hree physical aack ypes: (a) a passive eavesdropping aack, (b) a simple spoofing aack where a malicious acuaor blindly injecs a disrupive signal, and (c) an advanced spoofing aack where an adversary uses a sensor o measure he original signal and an acuaor o acively cancel he original signal and injec a malicious one. Acuaor Shield probe response Measured Eniy Malicious Malicious Acuaor o leverage only he physical delays addressed above, bu addiional compuaional delays would make i even easier o deec he presence of an aack. 2. Physical Aack Types for s Aacks can be classified as eiher passive (eavesdropping) or acive (spoofing). While we consider only physical/analog aacks in accordance wih assumpions A1 A4, he passiviy of an aack is decided by wheher or no he aacker is manipulaing (or spoofing) he physical signal or merely lisening o i. Acive aacks hemselves can be classified once more ino simple spoofing or advanced spoofing aacks. In shor, physical sensor aacks in accordance wih assumpions A1 A4 can be broadly divided ino hree caegories (Types): T1 Eavesdropping Aacks: In an eavesdropping aack, an adversary uses a malicious sensor in order o lisen o he acive sensor s communicaion wih he measured eniy (Figure 3a). T2 Simple Spoofing Aacks: In a simple spoofing aack, an adversary uses a malicious acuaor o blindly injec a malicious signal in order o aler he signal observed by he sensor. These aacks are simple in ha he malicious signal is no a funcion of he original, rue signal (Figure 3b). T3 Advanced Spoofing Aacks In an advanced spoofing aack, an adversary uses a sensor in order o gain full knowledge of he original signal and hen uses a malicious acuaor o injec a malicious signal accordingly. This enables an aacker o suppress he original signal or oherwise aler i in addiion o injecing a malicious signal (Figure 3c). We argue ha hese aack ypes span all possible modes of aacks ha abide by Assumpions A1 A4 wih hose goals oulined in G1 G3. For example, jamming or Denial of service (DoS) aacks falls in caegory T2 where he aacker s acuaor is used o blindly generae high ampliude, wide bandwidh signals o inerfere wih he physical signal before i reaches he sensors; replay aacks fall in eiher caegory T2 or T3 based on wheher he aacker is blindly replaying a physical signal or desrucing he original physical signal before insering he replay signal; spoofing aacks like hose demonsraed in [19] fall in caegory T2; and aacks described in [33] fall wihin boh T2 and T3. A firs glance, aacks of ype T1 may no seem imporan especially if he sensor under aack measures a physical signal ha is publicly accessible (e.g., room emperaure, car speed, ec.). In such cases, an adversary can measure he same physical signal wihou he need o lisen o he ineracion beween he acive sensor and he environmen. However, his may no always be he case. For example, an aacker migh measure magneic waves during an exchange beween an RFID reader and an RFID ag, learning poenially sensiive informaion abou he ag. These aacks are passive, meaning ha he aacker does no injec any energy ino he sysem. Secions describes mehods for deecing aack ypes T2 and T3, leaving aack ype T1 for laer discussion in Secion PYCRA AUTHENTICATION SCHEME The core concep behind PyCRA is ha of physical challengeresponse auhenicaion. In radiional challenge-response auhenicaion schemes, one pary requires anoher pary o prove heir rusworhiness by correcly answering a quesion or challenge. This challenge-response pair could be a simple password query, a random challenge o a known hash funcion, or oher similar mechanisms. In he proposed physical challenge-response auhenicaion, he challenge comes in he form of a physical simulus placed on he environmen by an acive sensor. Unlike radiional schemes, he proposed physical challenge operaes in he analog domain and is designed so ha an adversary canno issue he correc response because of immuable physical consrains raher han compuaional or combinaorial challenges. We begin by modeling he problem of deecing physical sensor aacks as an auhenicaion problem. To draw his analogy, le us consider he communicaion sysem shown in Figure 4a. This figure shows wo paries : (1) an acive sensor composed of acuaion and sensing subsysems and (2) he measured eniy which responds o signals emied by he acuaor conained wihin he acive sensor. The firs pary he acive sensor is responsible for iniiaing he communicaion by generaing some physical signal such as a magneic, acousic, or opical wave. The second pary he measured eniy responds o his communicaion by modulaing his signal and reflecing i back o he sensing subsysem of he acive sensor. Wih his analogy in mind, he problem of deecing physical aacks can be posed as ha of ensuring ha he message seen by he sensor has originaed from a rused pary (he rue eniy o be measured). This is akin o ideniy auhenicaion in he he lieraure of compuer securiy bu applied o he analog domain. 3.1 Simple PyCRA Aack Deecor Using he communicaion analogy shown in Figure 4a and recalling ha we are ineresed only in acive sensors as described in Secion 2.1, we noice ha he measured eniy, as a paricipaing pary in his communicaion, is sricly passive, i.e. i canno iniiae communicaion; i responds only when he sensor generaes an appropriae physical signal. PyCRA explois his passiviy in order o faciliae he deecion of aacks. Wihou PyCRA, an acive sensor s acuaor would probe he measured eniy in a normal fashion using a deermin- 17

5 Acive u() Acuaor A () Measured Eniy Acive u() Acuaor B() Measured Eniy Acive u() Acuaor B() Measured Eniy y() a() y() a() y() a() (a) (b) (c) Figure 4: An illusraion of he PyCRA archiecure and aack deecion scheme: (a) During normal operaion, he acive sensor generaes a signal A (). This signal passes hrough environmenal dynamics and is refleced back o he sensor as y(); (b) Using he proposed PyCRA scheme, he sensor generaes a modulaed signal B(). If here is no aack presen, he refleced signal diminishes if he acive sensor s acuaor is driven o zero; (c) Using he proposed PyCRA scheme while he sensor is under aack (by signal a()), a malicious signal is deeced during he period when he acuaor is disabled. isic signal denoed by A (). We embed in his signal a physical challenge hrough pseudo-random binary modulaion of he form: B() = u()a (), u() {,1} (1) where u() is he binary modulaion erm and B() is he modulaed oupu of he acuaor. The oupu of he acive sensor is denoed by y() as shown in Figure 4. In he absence of an aacker and from he passiviy of he measured eniy, seing u() = (and consequenly B() = ) a ime challenge will cause y() o go o zero. Poenial aackers mus acively emi a signal a() o overpower or mask y() (Goals G2 G3). A naïve aacker migh coninue o emi his signal even when B() = as shown in Figure 4c. In his case, he aack can be easily deeced, since any nonzero y() while u() = can be aribued o he exisence of an aacker. More advanced aackers migh aemp o conceal heir aacks when hey sense he absence of B() as in Goal G1. Due o Assumpion A3, an aacker could drive a() o zero only afer a delay of τ aack, where τ aack τ physical limi > is he unavoidable physical delay inheren in he aacker s hardware. Therefore, he mechanism described above can sill deec he presence of an aack wihin his unavoidable ime delay. Furhermore, an aacker canno learn and compensae for his inheren delay preempively due o he randomness of he modulaion erm u(). Again, any nonzero y() sensed while u() = can be aribued o he exisence of an aacker. The simple PyCRA aack deecor can be summarized as follows: [Sep 1] Selec a random ime, challenge [Sep 2] Issue a physical challenge by seing u( challenge ) = [Sep 3] If y( challenge ) >, declare an aack Noe ha he previous process needs o happen wihin small amoun of ime (e.g., in he order of milliseconds) such ha i does no affec he normal operaion of he sysem. 3.2 The Confusion Phase Every physical signal is subjec o random perurbaions known as noise. A fundamenal characerisic of his noise is he signal o noise raio (SNR). This SNR deermines he abiliy of any sensor o disinguish beween changes in a signal of ineres and he random noise. As wih he physical delay, his SNR is fundamenal, and i is never equal o zero. As a resul, if a signal is wihin he noise floor (less han he noise ampliude), i is fundamenally impossible o deec any change in he physical signal [36]. As wih he physical ime delay τ aack, we use his fundamenal limi in order o enhance PyCRA and inroduce addiional securiy. To do so, we modify he physical challenge by inroducing an inermediae sep beween he acive phase (e.g., u() = 1) and he silen phase (e.g., u() = ) called he confusion phase. In his phase, he acive sensor uses is acuaor o generae a signal u() ha is small enough o barely exceed he noise level. Nex, we wai in his confusion phase for a random ime con f usion before enering he silen ime. This process is summarized in Figure. Recall ha one of he aacker s goals is o remain sealhy (Goal G1). If he aacker is unable o insananeously deec he changes in he physical challenge, he or she will reveal hemselves. Due o he exisence of noise, no aacker wheher using sofware or hardware o couner he physical challenges issued by PyCRA can insananeously deec he change in he physical challenge. Tha is, here always exiss a non-zero probabiliy of he aacker missing he changes in he physical challenge. In Secion 4, we deail a heoreical resul ha explains he relaionship beween he ampliude of he physical challenge wihin he confusion phase and he probabiliy ha he aacker will fail o deec changes in he physical challenge. 3.3 Effec of Physical Delays a he As wih he aacker, he acuaor used by he acive sensor iself suffers from physical delays. This means ha when PyCRA issues a physical challenge, he acuaor oupu does no ransiion immediaely. Apparenly, if he physical delay in he acive sensor is greaer han τ aack, hen an adversary can conceal his signal. To couner his, PyCRA consrucs a mahemaical model for he sensor ha is used in real ime o predic and eliminae he effecs of he acive sensor s physics. By calculaing he residual beween he expeced oupu and he measured oupu, PyCRA can sill deec he exisence of an aack even if he sensor s dynamics are slower han hose of he aacker. 3.4 χ 2 PyCRA Aack Deecor If we obain an accurae model of he sensor s acuaor dynamics, hen we can remove is effecs from he measured response, ensuring ha any residual energy measured while u() = belongs o an exernal source such as an aacker Obaining he Model To compensae for he acuaor dynamics, we firs need o acquire an accurae model ha capures he underlying physics of he acive sensor. Below we model he acive sensor using he generic nonlinear sae updae of he form: x( + 1) = f (x(),u()) + w() (2) y() = h(x()) + v() (3) where x() R n is he acive sensor sae a ime N (e.g., he elecrical curren and volages inside he sensor a ime ), u() R is he modulaion inpu o he sensor, he funcion f : R n R R n 18

6 he can reac o hese challenges. In his secion, we show a heoreical resul ha allows PyCRA o increase he probabiliy of an aacker failing o deec he changes in he physical-challenge by correcly designing he confusing phase (discussed in Secion 3.2) and hence increase he probabiliy of deecing he aack. T HEOREM 1. Consider an aacker aemping o deec changes in a physical challenge signal wih mis-deecion probabiliy α. For any sraegy he aacker chooses, and because of he SNR exiss a any sensor, he probabiliy of he aacker having a consan deecion delay τ > is bounded away from zero, i.e., wih high probabiliy he aacker will deec a change and urn off his signal only afer ime T afer he beginning of he confusion period. In addiion, decreasing he ampliude of he signal emied by he acive sensor during he confusion period by a facor of β > 1 increases he delay τ by a facor of β 2. P ROOF S KETCH. We base he proof on he resuls repored in [36] on change poin deecion which measure fundamenal limis on checking changes in noisy signal. In he change poin deecion seing, he false alarm probabiliy is analogous o he even where he aacker swiches off his signal before he beginning of he silen period. Delay in [36] is defined as he ime ha elapses from he change poin unil he change is deeced. When α 1, he false alarm probabiliy induces a probabiliy which is proporional o α for he even ha change is deeced wihin a ime inerval shorer han T (a consan independen of α). Decreasing he ampliude of he signal acuaed by he acive sensor during he confusion period by a facor of β leads o a decrease in SNR by a facor of β 2. Based on his relaion, he aacker has o increase he delay by a facor of β 2 in order o mainain false alarm probabiliy α. Figure : acuaor oupu (op) wih confusion and silence phases and he corresponding raw signal (boom) wih an aack. is a model describing how he physical quaniies of he sensor evolve over ime, and he funcion h : Rn R models he sensor measuremen physics. Such models can be eiher derived from firs principles [9, 2, 1] or hrough experimenal sudies [23, 2]. Addiionally, hese models are used o design he sensors hemselves and are ypically known o he sensor manufacurers. Finally, since no mahemaical model can capure he rue sysem behavior exacly, he erm w() Rn represens he mismach beween he rue sensor and he mahemaical model while v() models he noise in he sensor measuremens. χ 2 Deecor We use he dynamical model of he sensor (Equaions (2) and (3)) in designing a χ 2 deecor o deec he exisence of an aacker. χ 2 deecors appear in he lieraure of auomaic conrol, where hey are used in designing faul oleran sysems [26, 28, 2, 38]. The χ 2 deecor works as follows: [Sep 1] Selec random imes, challenge and con f usion. [Sep 2] Issue a physical challenge by enering he confusion phase a ime challenge and hen ener he silen phase a ime challenge + con f usion. [Sep 3] Residual Calculaion: Here we use Equaions (2) and (3) o calculae an esimae for he curren sensor sae x() ˆ and he prediced oupu y() ˆ = h(x()). ˆ This operaion is iniiaed a challenge + con f usion when u() ransiions o he acuaor silence ime and erminaes once u() ransiions back o one, signaling he end of acuaor silence. The model represened by Equaions (2) and (3) describes he oupu of he sensor when he aack is equal o zero. Therefore, he residual1 beween he measured oupu and he prediced oupu, z() = y() y(), ˆ corresponds o boh he aack signal as well as he environmenal dynamics during he ime inerval before u() drops o. For each segmen of lengh T where u() =, we calculae he norm of he residual z() as: g() = 1 T z2 (τ). Magneic encoders are acive sensors used in a wide array of indusrial, roboics, aerospace, and auomoive applicaions. The goal of an encoder is o measure he angular velociy or posiion of a gear or wheel in order o provide feedback o a moor conroller. The operaion of hese sysems depends heavily on he accuracy and imeliness of he individual encoders. This secion describes he basic operaion of magneic encoders in paricular and he ypes of aacks ha can be mouned agains hem as well as how PyCRA can be used o provide securiy for hem..1 (4) Magneic Encoders Magneic encoders rely on magneic variaions o measure he angular velociy of a gear or wheel and are ofen designed o handle dus, mud, rain, and exreme emperaures wihou failing. The goal of each encoder is o provide a signal whose frequency corresponds o he speed of a gear. These signals are condiioned and passed o a moor conroller uni which deecs if any correcive acions need o be aken. Typical magneic encoders operae by generaing a magneic field in he presence of a roaing ferromagneic one ring or one wheel. This ring has a number of eeh on is edge so ha he refleced magneic wave as observed by he encoder varies over ime as a (noisy) sinusoidal wave. By measuring he frequency of his refleced signal over ime, each sensor and consequenly he moor conroller is able o infer he angular velociy of any given gear, wheel, or moor as illusraed in Figure 6. Aacks on magneic encoders have been sudied in [33] in he conex of Ani-lock Braking Sysems in auomoive vehicles. Boh τ= T +1 [Sep 4] Deecion Alarm: Once calculaed, we compare he χ 2 residual g() agains a pre-compued alarm hreshold α. This alarm hreshold is chosen based on he noise v(). Whenever he condiion g() > α is saisfied, he sensor declares ha an aacker has been deeced. 4. CASE STUDY (1): DETECTING ACTIVE SPOOFING ATTACKS FOR MAGNETIC ENCODERS THEORETICAL GUARANTEES As discussed before, PyCRA is based on he concep ha physics impose fundamenal and immuable consrains on how quickly an aacker can deec changes in he physical-challenge and how fas 1 The name of he Chi-squared (χ 2 ) deecor follows from he fac ha, in he case of no aack, he residual z() is a Gaussian random variable, and hence is square g() is a χ 2 disribued random variable. 19

7 Encoder Moor Conroller Gear Figure 6: Flow diagram for a ypical magneic encoder: The signal begins as a refleced magneic wave from a gear. This signal is capured by a pick-up coil or Hall Effec sensor, condiioned ino a clean square wave, and finally ranslaed ino an angular velociy. simple spoofing [T2] and advanced spoofing [T3] aacks are shown o influence he vehicle sabiliy. In his case sudy, we show how PyCRA can deec he exisence of such aacks..2 The PyCRA-secured Magneic Encoder Physically, he proposed secure magneic encoder sensor consiss of wo main pars: (i) he fron-end conaining he acuaor and pickup coils responsible for boh probing he roaing one ring and measuring he response, and (ii) he processing backend. Figure 7 shows he fron-end of he sensor used in our evaluaion. The acuaor coil depiced is much larger han would be required in a commercial produc, because i consiss of a magneic core and a hand-wound high-gauge wire. The following is an overview of he main blocks of he sensor..2.1 Acuaor Coil The main componen required for he secure sensor is he acuaor coil. In his work, we use an insulaed copper wire wrapped around a ferromagneic core and driven using a power amplifier..2.2 Pickup and Filering The pickup (measuremen) coil is wrapped around he same ferromagneic core used for he acuaor coil. In order o reduce he effec of noise from oher EMI sources wihin he vehicle body, he oupu of he pickup coil is conneced o a differenial amplifier wih high common-mode rejecion. The oupu of his differenial amplifier is conneced o he digial processing backend. Anoher securiy concern of he magneic encoder is he wires connecing he coils o he digial backend. These wires pose a poenial vulnerabiliy, as an aacker can cu hem and connec his aack module direcly. However, such aacks are already accouned for in many sysems as addressed in Assumpion A Processing Elemens The secure sensor requires enough processing power o perform he necessary compuaions in real-ime. The DSP calculaions ake place on a high power ARM Corex (M4 STM32F47) processor, which has ample floaing poin suppor. We do no consider any power consumpion issues in our design..3 Obaining he Model The dynamics of he sensor (including he acuaor, high gain curren amplifier, sensors, and he signal condiioning circui) are idenified using sandard sysem idenificaion mehods [23]. Tha is, we applied four differen pseudo random binary sequences (PRBS) o he sysem, colleced he oupu, and hen applied subspace sysem idenificaion echniques in order o build models of increasing complexiy [23]. Finally we used boh whieness ess and correlaion ess o assess he qualiy of he obained model [2]. In order o validae he model, we generaed a random sequence similar o hose used in he real implemenaion of he sensor. We fed he same inpu o boh he sensor and he model and recorded he error. Acuaor Coil Pickup Coil Gear Figure 7: PyCRA encoder acuaor coil, sensor, and gear seup. Experimens show ha he model is reasonably accurae wih an error in he range of milli-vols..4 Tesbed In order o es he PyCRA-secured magneic encoder, we consruced a esbed consising of he proposed secure sensor aached o a Mazda Rx7 one ring. The one ring is aached o a DC moor which simulaes a roaing wheel. An addiional coil is added o simulae he effec of an aacker. The aacker coil is also conrolled by a high gain amplifier conrolled hrough a real-ime xpc Targe sysem conneced o MATLAB. A Mazda RX7 magneic encoder sensor is also aached o he same one ring in order o provide ground ruh. The oupu of his sensor is conneced o a MAX9926U evaluaion ki which includes an inerface capable of convering he raw sinusoidal wave ino he encoded square wave as shown in Figure 6. The oupu of he proposed secure sensor as well as he oupu of he MAX9926U is moniored by he same real-ime xpc Targe for comparison.. Calibraion agains naural variaions modeling is usually done in a conrolled environmen. However, once he sensor is placed in a esbed, muliple naural variaions, mechanical asymmeries, and oher environmenal facors degrade he accuracy of such models. To accoun for hese variaions, we use a simple learning mechanism o esimae he noise level in he measuremens and he deviaion beween he expeced oupus (as per he model) and he acual oupus. Once hese parameers are learned, we can se he alarm hreshold accordingly. Resuls can be furher improved by considering online idenificaion-and-calibraion of he sensor model..6 Aack Deecion for Magneic Encoders We begin wih a simple spoofing aack [T2] in which an aacker injecs a sinusoidal wave of varying frequency. Spoofing aacks of his naure aemp o overpower he rue frequency of he sysem and force he sensor o rack he false frequency (mirroring he simplisic spoofing aack in [13]). In his experimen he original one ring frequency is fixed a 71 Hz, and he frequency of he aacking coil increases linearly from 6 Hz o jus over 4 Hz. As per our aacker model in Secion 2, we assume ha he aacker aemps o conceal his or her presence (Adversarial goal [G1]). This means ha he adversary will be able o deec when he acuaor coil is urned off and will, afer some ime τ aack, emporarily hal he aack. The sealhiness of he aacker necessiaes ha he PyCRA deecion scheme have high accuracy even when he aacker is quick 11

8 True Posiive Rae Deecion F 1 Score ms. ms.2.2 ms ms False Posiive Rae (a) Fs = 3 khz.4 Fs = 2 khz Fs = 1 khz Aacker Physical Delay, τ aack (ms) (b) Figure 8: Resuls from 3 experimens showing (a) he accuracy of aack deecion for a simple spoofing aack wih sampling rae F s = 1 khz and a range of τ aack, and (b) aack deecion accuracy as a funcion of τ aack for several sampling raes, F s. o reac. We evaluaed he PyCRA deecion scheme across a range of τ aack values, χ 2 deecion hresholds (α), and sampling frequencies (F s ). Noe ha in order o simulae an aacker wih ms physical delays (which is physically impossible), we gave he aacker access o he random signal generaed by PyCRA so ha he aacker can sar shuing down his acuaors before PyCRA generaes he physical challenge. In oal, we conduced over 3 experimens on our experimenal esbed o validae he robusness of he proposed securiy scheme. The resuling accuracy wih F s = 1 khz is depiced by he ROC 2 curves in Figure 8a for a range of α. From his figure i is clear ha beween τ aack = and 7 µs is all ha is necessary for PyCRA o accuraely disinguish aacked signals from normal signals, if α is chosen appropriaely. Wih α se o a predeermined value, we can vary F s as shown in Figure 8b 3. These resuls show ha increasing F s from 1 khz o 3 khz reduces required ime for deecion o beween τ aack = 1 and 2 µs. Repeaing hese experimens for he advanced spoofing aack [T3] yields similar resuls. In fac, here is no fundamenal difference beween he wo in erms of aack deecion; his is governed by he dynamics of he aacker s acuaor raher han he naure of he aack iself. I is imporan o evaluae his deecion accuracy (which is our securiy guaranee) in erms of he physical delay propery τ aack of he aacker model. In pracice, he sae-of-he-ar in low-dimension, high Q-facor hardware ha provide enough power o carry ou a spoofing aack will have τ 2µs 4. From Figure 8b i is apparen ha PyCRA has good performance for his range of pracical physical delays. Moreover, he resuls we have shown hus far use a relaively low sampling frequency (high end micro conrollers can operae in he range of 2 khz). As illusraed by Figure 8b, higher sampling 2 A Receiver Operaing Characerisic (ROC) is a visual aid for evaluaing he accuracy of binary classifiers in erms of boh rue and false posiive raes. 3 The F 1 score is a saisical measure of a binary classifier ha measures he classifier accuracy in erms of precision and recall. 4 These values were obained by surveying a range of sae-of-hear, commercially available componens. Proxmark3 Oscilloscope RFID Reader Sniffer Tag (a) (b) Figure 9: The schemaic used in he RFID eavesdropping case sudy (a) and corresponding hardware seup (b). The seup conains wo low frequency anennas (one for he RFID reader and one for he eavesdropper) along wih a Proxmark3 board running he PyCRA deecion algorihm. The analog signal is also capured by a digial oscilloscope for visualizaion raes resul in reduced aack deecion imes. However, using low sampling frequencies in our case sudy serves o illusrae he efficiency of he proposed deecion mechanism. 6. CASE STUDY (2): DETECTION OF PAS- SIVE EAVESDROPPING ON RFID In his secion, we discuss deecion of passive eavesdropping aacks on acive sensors. In his scenario, an adversary lisens or records he same physical signals capured by he sensor. Indeed his ype of aack saisfies assumpions A1 A3 described in Secion 3 and hence i will be useful o exend PyCRA o such cases. 6.1 Passive Eavesdropping on RFID In his secion, we use radio-frequency idenificaion (RFID) as an example where successful passive aacks can have severe consequences. RFID sysems are commonly used o conrol access o physical places and o rack he locaion of cerain objecs. An RFID sysem consiss of wo pars: a reader and a ag. The RFID reader coninuously sends a magneic signal ha probes for exising ags in he near proximiy. Once a ag eners he proximiy of he reader, he ag sars o send is unique idenifier o he reader by modulaing he magneic probe signal. RFID ags can be classified as eiher passive or acive based on he source of heir energy. While passive ags rely on he energy ransmied by an RFID reader in order o power heir elecronics, acive ags have heir own power supplies. As a resul, acive ags can be equipped wih compuaional plaforms ha run crypographic and securiy proocols o miigae cloning aacks []. On he oher hand, passive ags do no enjoy hese luxuries and herefore are more prone o cloning aacks. Cloning of passive RFID ags can ake place in one of wo forms. In he firs, he aacker uses a mobile RFID reader and aemps o place i near he RFID ag. The ag innocenly responds o he communicaion issued by he adversarial reader and sends is unique idenifier. The oher form of aack carried ou agains RFID sysems is o eavesdrop on he communicaion beween he ag and a legiimae reader. RFID proecive shields and blockers [27, 1] are commonly used as counermeasure o he firs form of cloning aacks discussed above. Unforunaely, hese shields are of no use 111

9 Volage [vol] 1 Volage [vol] 1 Volage [vol] Volage [vol] 1 Volage [vol] 1 Volage [vol] Volage [vol] 4 2 Alarm hreshold Volage [vol] 4 2 Alarm hreshold Volage [vol] 4 2 Alarm hreshold Normal response (no aack) Response under aack Residual = oupu - expeced Normal response (no aack) Response under aack Residual = oupu - expeced Normal response (no aack) Response under aack Residual = oupu - expeced (a) (b) (c) Figure 1: Resuls of applying PyCRA o deec he exisence of an eavesdropper in he proximiy of an RFID reader. (a) Resuls of using sandard 12KHz signal for deecion. (b) Resuls of using PyCRA when only an RFID ag is presen in he proximiy of he PyCRA-enabled reader, and (c) Resuls of using PyCRA in deecing eavesdropping when boh an RFID ag and an eavesdropper anenna are presen in he proximiy of he PyCRA-enabled reader. Top figure shows he response o he physical challenges when no eavesdropper is placed in he proximiy of he RFID reader. The middle figure shows he response o he physical challenges when (a) an eavesdropper anenna (b) passive ag only (c) passive ag + eavesdropper anenna are placed in he proximiy of he reader. Finally, he boom figure shows he value of he residuals calculaed by PyCRA along wih he alarm hreshold. agains he second ype of aacks, because he user is obliged o remove he proecive shield before using he ag wih he legiimae RFID reader, a which ime an adversary can successfully eavesdrop. To carry ou an eavesdropping aack, a sniffing device mus be placed in close proximiy o he RFID reader so ha i can measure he elecromagneic waves ransmied by he reader and refleced by he ag. In he following resuls, we show how PyCRA is able o deec he exisence of such an aack, allowing he reader o disable communicaion wih he ag before revealing privae informaion. 6.2 Using PyCRA o Deec Eavesdroppers Recall from he physics of elecromagneic waves ha anennas placed wihin close proximiy will always affec each oher s elecrical signals o some degree. This is a fundamenal law in physics known as magneic coupling [2] and is used in he design of RFID. Noe ha, similar o he physical delays, he magneic coupling assumpion is a fundamenal limiaion ha he aacker canno overcome. Hence, we can use he PyCRA deecion algorihm oulined in Secion 3 by compuing he residual beween he model (which assumes magneic coupling only wih he RFID ag) and he sensor oupu which suffers from magneic coupling wih he eavesdropping anenna. This is shown in he experimenal resuls in he nex subsecion. 6.3 Hardware Seup and Real-ime Resuls Figure 9 shows he hardware seup used o carry ou his casesudy. In his seup, wo idenical low frequency RFID anennas are used. The firs RFID anenna is used as he legiimae RFID reader while he second is used o eavesdrop. We buil a PyCRA-enabled RFID reader on op of he open source RFID Proxmark3 board, adding random modulaion o he probing signal and consrucing he appropriae models as oulined in Secion 3. Figure 1a (op) shows he received elecromagneic wave of an RFID reader operaing in he classical operaional mode. In his mode, he RFID reader generaes he sandard 12KHz sinusoidal wave ha is used o communicae wih he RFID ag. Figure 1a (middle) shows he resuling elecromagneic wave when an eavesdropper uses an idenical anenna o lisen. In his case i is hard o disinguish beween he wo waves and hence hard o deec he exisence of an eavesdropper. This is beer illusraed by he residual beween he wo waves as shown by he residual in Figure 1a (boom). On he oher hand, Figures 1b and 1c show he resul of he same case sudy when PyCRA is used wih and wihou an eavesdropper presen, respecively. In his mode, he PyCRA algorihm occasionally hals he normal operaion of he RFID reader and swiches o deecion mode. In his mode, PyCRA selecs randomized periods of ime o issue he physical challenges by swiching he RFID anenna from on o off and from off o on. In order o selec an appropriae alarm hreshold, we firs sudy he effec of he magneic coupling beween he reader and he ag in he absence of an eavesdropper. This is shown in Figure 1b where he alarm hreshold is chosen such ha no alarm is riggered when he effec of he magneic coupling he residual beween he no ag response (op) and he response wih a ag (middle) is wihin he accepable range. This accepable residual range allows for coupling induced by he ag only. Any increase on op of his allowable hreshold is hen aribued o he exisence of an excessive magneic coupling due o an eavesdropper. Figure 1c (middle) shows he response o he same se of physical challenges when he aacker places an eavesdropping anenna in he proximiy of he RFID reader while he ag is also presen. Thanks o he physical challenge produced by PyCRA, he magneic coupling produced by he eavesdropper anenna is now easily deeced. This can be shown in Figure 1c (boom) which shows he residuals beween he expeced oupu and he measured signal exceeding he alarm hreshold. We recorded over 1 hour of RFID measuremens wih varying disances of he malicious eavesdropping anenna. Of hose experimens where he aacker was close enough o observe he RFID communicaion, PyCRA successfully deeced he exisence of an 112