Secure Data Timestamping in Synchronization-Free LoRaWAN

05/05/2019 ∙ by Chaojie Gu, et al. ∙ Nanyang Technological University Peking University 0

Low-power wide-area network technologies such as LoRaWAN are important for achieving ubiquitous connectivity required by the Internet of Things. Due to limited bandwidth, LoRaWAN is primarily for applications of collecting low-rate monitoring data from geographically distributed sensors. In these applications, sensor data timestamping is often a critical system function. This paper considers a synchronization-free approach of timestamping the uplink data at the LoRaWAN gateway, which can give milliseconds accuracy. Its key advantages are simplicity and no extra overhead, commensurate with the scarce communication resources of LoRaWAN. However, we show that this low-overhead approach is susceptible to a frame delay attack that can be implemented by a combination of stealthy jamming and delayed replay. To address this threat, we propose a SoftLoRa gateway design that integrates a commodity LoRaWAN gateway with a low-power software-defined radio receiver to track the inherent frequency biases of LoRaWAN end devices. With a set of efficient signal processing algorithms that are designed based on LoRaWAN's modulation method, our frequency bias estimation achieves a resolution of 0.14 parts-per-million (ppm) of the channel's central frequency. This resolution is sufficient to detect the attack that introduces an additional frequency bias of one or more ppm. We evaluate our approach in various indoor and outdoor environments. In summary, this paper presents an attack-aware and low-overhead approach to timestamping the data generated by LoRaWAN end devices.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 7

page 11

page 12

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1. Introduction

Low-power wide-area networks (LPWANs) enable direct wireless interconnections among end devices and gateways in geographic areas of up to tens of square kilometers (Raza et al., 2017). It will increase the network connectivity as a defining characteristic of the Internet of Things (IoT). Among various LPWAN technologies (including NB-IoT and Sigfox), LoRaWAN (Sornin et al., 2016), which is an open data link layer specification based on the LoRa physical layer technique, offers the advantages of using license-free ISM bands (e.g., EU and US ), low costs for end devices, and independence from managed infrastructures (e.g., cellular networks).

LoRaWAN is promising for the applications of collecting low-rate monitoring data from geographically distributed sensors, such as utility meters, environment sensors, roadway detectors, industrial IoT measurement devices, and etc. Most of these real-world monitoring applications require the sensing data to have timestamps in the global time, though the timestamps do not have to be highly accurate such as microseconds level. For instance, in both indoor and outdoor environment condition monitoring, seconds accuracy for sensor data timestamps will be sufficient due to the slow dynamics of the environment condition. Second-accurate timestamps for the traffic data generated by roadway detectors can be used to reconstruct real-time traffic maps well. In a range of industrial monitoring applications such as oil pipeline monitoring (Byers, [n. d.]), milliseconds accuracy is sufficient.

Conventionally, to perform the data timestamping in a wireless sensor network (WSN) for data collection, the clocks of the sensor nodes need to be synchronized, such that the nodes can timestamp their data once generated. To achieve clock synchronization, each WSN node can be equipped with a GPS receiver for accessing the global time. However, GPS receivers consume excessive power and may not work in indoor environments. Thus, various WSN clock synchronization protocols based on message exchanges have been developed. Different from the above synchronization-based approach, the synchronization-free approach uses the gateway with a globally synchronized clock to timestamp the data upon the arrival of the corresponding network packet. However, in multi-hop WSNs, this synchronization-free approach may perform unsatisfactorily, because the data delivery on each hop may have uncertain delays due to various factors such as channel contention among nodes.

Differently, the synchronization-free approach is desirable for uplink data timestamping in LoRaWANs. Reasons are two-fold. First, in contrast to the multi-hop WSNs, LoRaWANs adopt a one-hop gateway-centered star topology that is free of the issue of hop-wise uncertain delays. Specifically, as the radio signal propagation times from the end devices to the gateway are generally in microseconds, the LoRaWAN frame arrival time instant can well represent the time instant that the frame leaves the end device. As a result, timestamping the uplink data at the gateway can meet the timestamping accuracy requirements of many real-world applications. Second, if the synchronization-based approach is adopted otherwise, the task of keeping the end devices’ clocks synchronized and the inclusion of timestamps in the LoRaWAN data frames will introduce considerable overhead to the narrowband LoRaWANs. Given the above reasons, the synchronization-free approach of timestamping the uplink data at the gateway is lightweight, implementation-friendly, and efficient due to its simplicity.

Despite the prospect of the synchronization-free approach owing to its low overhead, in this paper, we take an adversary perspective to examine this approach under the LoRaWAN context and seek to improve its security. This is because that wrong timestamping can also lead to undesirable consequences. The long-range communication capability of LoRaWAN, though increasing connectivity, renders the communications susceptible to wireless attacks that can be launched from remote and hidden sites. The LoRaWAN specifications only define conventional frame confidentiality, integrity, and device authentication measures. These conventional security measures may be inadequate to protect the network from wireless attacks on the physical (PHY) layer of LoRaWAN.

In this paper, we consider a basic threat of frame delay attack described in the RFC 7384 (Mizrahi, 2014) that will subvert the synchronization-free uplink data timestamping approach. Our experiments based on a commodity LoRaWAN platform show that there is a time window of tens of milliseconds (ms) after the onset of a legitimate frame transmission for implementing stealthy jamming. The jammed victim gateway cannot decode any frame and raises no alerts to the operating system (OS). The lengthy time window makes the stealthy jamming easily achievable using commodity LoRa devices. Based on the above, we have implemented the frame delay attack through a combination of stealthy jamming and frame replay that can introduce arbitrary delays to the deliveries of LoRaWAN frames.111Two computer science undergraduate students have also implemented the attack with light instructions from us and documented their results in their dissertations (Zhu, 2017; Tay, 2018). With the information provided in this paper, skillful attackers can easily re-implement the attack. This will subvert the synchronization-free data timestamping approach that assumes near-zero signal propagation times. As the attack does not breach frame integrity, it cannot be solved by cryptographic protection and conventional security measures such as frame counting.

To address the above attack, we explore the LoRaWAN transmitters’ traits that can be extracted from their transmitted signals. A recent study (Eletreby et al., 2017) exploited the LoRaWAN transmitters’ distinct frequency biases (FBs) in generating the chirps to disentangle colliding frames. The biases are mainly caused by the manufacturing imperfections of the transmitters’ internal oscillators. Inspired by this, we inquire whether the replay step of the frame delay attack introduces extra detectable FBs. Different from (Eletreby et al., 2017) that only needs a coarse-grained analysis to detect multiple peaks in the frequency domain, we will need high-precision FB estimation to detect a small extra FB introduced by the malicious replayer.

To this end, we propose a SoftLoRa gateway that integrates a low-power software-defined radio (SDR) receiver (rtl, 2018) with a commodity LoRaWAN gateway to capture and analyze the received radio signals. Compared with an alternative solution of using a single full-fledged SDR transceiver such as USRP to demodulate and analyze the received signals in software, our SoftLoRa gateway is advantageous in its efficient hard-speed demodulation while its SDR receiver is used for defense only. Based on LoRa’s Chirp Spread Spectrum (CSS) modulation method, we develop a set of signal processing algorithms for SoftLoRa to estimate the FB. From our experiments with 16 LoRaWAN end devices, we show that (i) with a received signal-to-noise ratio (SNR) of down to , SoftLoRa achieves a resolution of in estimating the transmitter’s FB, which is just 0.14 parts-per-million (ppm) of the channel’s central frequency of ; (ii) the frame replay by a USRP introduces an additional FB of at least (i.e., 0.62 ppm), exceeding SoftLoRa’s FB estimation resolution of 0.14 ppm. Thus, SoftLoRa can track FB to detect the replay step of the frame delay attack. Note that the detection does not require uniqueness or distinctiveness of LoRaWAN end devices’ FBs, because it detects the FB changes caused by the replay, rather than identifies the transmitter.

In this paper, we make the following contributions:

  • We implement a stealthy frame delay attack with quantified parameter ranges against a commodity LoRaWAN platform, alerting to the insecurity of any system functions that rely on timely delivery of frames such as the gateway’s uplink data timestamping.

  • We design time-domain signal processing algorithms for accurately estimating LoRaWAN end devices’ FBs. The FB estimation requires accurate timestamping of the signal arrival. We achieve microseconds signal timestamping accuracy, improving our understanding on the timestamping accuracy for narrowband LoRaWAN signals and also echoing the results in (Ramirez et al., 2019) that are obtained using a different approach.

  • With the accurate FB estimation, our SoftLoRa gateway can reliably detect the frame delay attack. The SoftLoRa gateway enables an attack-aware lightweight approach to timestamping the data generated by LoRaWAN end devices that run little or even no code for timestamping. The simplicity, low overhead, and attack awareness are highly desired in real-world systems.

The rest of this paper is organized as follows. §2 reviews related work. §3 details the synchronization-free data timestamping in LoRaWANs. §4 studies the frame delay attack. §5 presents the SoftLoRa gateway design. §6 develops the microseconds-accurate signal timestamping needed by FB estimation. §7 studies LoRa’s FB and uses it to counteract the frame delay attack. §8 presents the results of the experiments in real environments. §9 concludes this paper.

2. Related Work

The communication performance of LoRaWAN has received increasing research. A theoretical capacity analysis for LoRaWAN is presented in (Mikhaylov et al., 2016). LoRaWAN’s communication performance is profiled via field measurements (Petrić et al., 2016; Toldov et al., 2016; Marcelis et al., 2017). Marcelis et al. (Marcelis et al., 2017) propose a coding scheme for data recovery. The Choir system (Eletreby et al., 2017) exploits the diverse FBs of the LoRaWAN end devices to decode colliding frames from different end devices. However, it does not develop an FB estimation algorithm. The Charm system (Dongare et al., 2018) exploits coherent combining to decode a frame from the weak signals received by multiple geographically distributed LoRaWAN gateways. It allows the LoRaWAN end device to use a lower transmission power. Several recent studies (Talla et al., 2017; Peng et al., 2018; Hessar et al., 2019; Varshney et al., 2017) have devised various backscatter designs for LoRa to reduce the power consumption of end devices. However, all these existing studies focus on understanding and improving the data communication performance of LoRaWAN (Mikhaylov et al., 2016; Petrić et al., 2016; Toldov et al., 2016; Marcelis et al., 2017, 2017; Eletreby et al., 2017; Dongare et al., 2018), or reducing power consumption via backscattering (Talla et al., 2017; Peng et al., 2018; Hessar et al., 2019; Varshney et al., 2017). None of them specifically addresses efficient data timestamping, a basic system function of many LoRaWAN-based systems.

LongShoT (Ramirez et al., 2019) is an approach to synchronize the LoRaWAN end devices with the gateway. Through low-level offline time profiling for a LoRaWAN radio chip (e.g., to measure the time delays between hardware interrupts and the chip’s power consumption rise), LongShoT achieves sub-50 microseconds accuracy. LongShoT is designed for the LoRaWAN systems requiring tight clock synchronization. Differently, we address data timestamping and focus on the less stringent but more commonly seen milliseconds or sub-second accuracy requirements. Moreover, as we will discuss in §3.2, the timestamping approach based on the prior clock synchronization will introduce considerable overhead. If highly accurate (e.g., microseconds accuracy) timestamping is not needed, it is wise to adopt the synchronization-free approach with the security enhancement presented in this paper.

Security of LoRaWAN has received limited research. In (Aras et al., 2017a), Aras et al. discuss several possible attacks against LoRa, including key compromise, frame replay, and jamming. The first two attacks need prior physical attacks such as memory extraction and node reset. Their jamming simply aims at subverting the victim receiver’s frame decoding. In this paper, we additionally examine the timing of the jamming such that the victim gateway does not alert the OS. In (Aras et al., 2017b), a selective jamming attack against certain receivers and/or certain application frames is studied. From our results in §4.3, the selective jamming in (Aras et al., 2017b) cannot be stealthy because it cannot start jamming until the frame header is decoded. As a result, the selective jamming will corrupt the payload, leading to integrity check failures and alerts. In (Robyns et al., 2017)

, Robyns et al. apply deep learning for LoRa transmitter identification based on the received baseband signal. Their approach can only identify the source transmitter as one of the considered transmitters that the trained deep model captures. It cannot be used to detect the malicious replayer that is in general out of the deep model.

3. Data Timestamping in LoRaWAN

3.1. LoRaWAN Primer

LoRa is a PHY layer technique that adopts a CSS modulation and works in ISM bands (e.g., US and EU ). LoRaWAN is an open data link specification based on LoRa. A LoRaWAN is a star network consisting of a number of end devices and a gateway that is often connected to the Internet. Gateways are often equipped with GPS receivers for time keeping. The transmission direction from the end device to the gateway is called uplink and the opposite is called downlink. LoRaWAN defines three classes for end devices, i.e., Class A, B and C. In Class A, each communication session must be initiated by an uplink transmission. There are two subsequent downlink windows. Class A end devices can sleep to save energy when there are no pending data to transmit. Class A adopts the ALOHA media access control protocol. Class B extends Class A with additional scheduled downlink windows. However, such scheduled downlink windows will require the end devices to have synchronized clocks, incurring considerable overhead as we will analyze shortly. Class C requires the end devices to be in the listening mode all the time. Clearly, Class C is not for low-power IoT objects. In this paper, we focus on the energy-efficient Class A, because it is supported by all commodity LoRaWAN platforms. To the best of our knowledge, no commodity LoRaWAN platforms have out-of-the-box support for Class B – the system developers will need to engineer the needed clock synchronization first.

3.2. Sync-Based vs. Sync-Free Timestamping

Data timestamping, i.e., to record the time of interest in terms of wall clock that is meaningful to the data, is a basic system function required by the monitoring data collection applications based on LoRaWAN. For a sensor measurement, the time of interest is the time instant when the measurement is taken by the end device. WSNs largely adopt the synchronization-based approach. Specifically, the clocks of the WSN nodes are synchronized to the global time using some clock synchronization protocol. Then, each WSN node can timestamp the data using its local clock. WSNs have to adopt this approach due primarily to that the multi-hop data deliveries from the WSN nodes to the gateway in general suffer uncertain time delays. Thus, although the clock synchronization introduces additional complexity to the system design, it has become a standard component for systems requiring data timestamping. However, in LoRaWANs with much less communication capacity due to their narrowband nature, the overhead of the clock synchronization cannot be ignored.

Now, we present an example to illustrate. Assume we have a microseconds or milliseconds accurate clock synchronization approach for LoRaWAN. Typical crystal oscillators found in microcontrollers and personal computers have drift rates of 30 to 50 ppm (Hao et al., 2011). Without loss of generality, we adopt 40 ppm for the following calculation. Under this drift rate, an end device will need 14 synchronization sessions per hour to ensure a sub- clock error. These synchronization sessions may represent a significant communication overhead for an end device. For instance, in Europe, a LoRaWAN end device adopting a spreading factor of 12 can only send 24 30-byte frames per hour to conform to the 1% duty cycle requirement specified by The European Telecommunications Standards Institute (Institute, [n. d.]). Although the synchronization information may be piggybacked to the data frames as in (Ramirez et al., 2019), a low-rate monitoring application may have to send the frames more frequently just to satisfy the time keeping requirement. In addition, the timestamps for data records also occupy frame payload space. For instance, if each frame with 30-byte payload contains a eight-byte timestamp (Ramirez et al., 2019) for the data in the frame, 27% of the effective bandwidth is used to convey timestamps. From the above example, we can see that the clock synchronization service and the data timestamps may consume a significant fraction of LoRaWAN’s communication capacity.

In this paper, we consider a synchronization-free approach. Specifically, a LoRaWAN end device still records the times of interest in terms of its unsynchronized clock. Right before sending a number of data records using a frame, the device replaces the the records’ times of interest in its local clock with their elapsed times up to the present, form the frame, and transmit it immediately. We assume that the buffer time from the generation to the transmission of the data records is short to ensure limited local clock drift and limited bits to represent the elapsed times. For instance, to enforce an upper bound of clock drift under a drift rate of 40 ppm, the buffer time needs to be within 4.1 minutes. As a result, 18 bits will be sufficient to represent an elapsed time with resolution. Since the one-hop signal propagation time from the end device to the gateway is negligible for millisecond-level systems, the gateway can easily reconstruct the global timestamps based on the frame arrival time and the elapsed times contained in the frame. Compared with the synchronization-based approach, this synchronization-free approach avoids the communication overhead and implementation complexity caused by clock synchronization. It can also reduce the frame payload use for time information (e.g., 18 bits for elapsed time versus 8 bytes for complete timestamp). In particular, if the end device can immediately transmit any newly generated sensor/event data, the elapsed time payload is even not needed.

The synchronization-free approach is lightweight and commensurate with LoRaWAN’s scarce communication resources. Its accuracy can be affected by (1) the delay from the end device’s application code requesting frame transmission to the actual emission of radio signal by the LoRaWAN chip and (2) the gateway’s accuracy in timestamping the radio signal arrival. An existing study (Gu et al., 2018) shows that these issues cause a sum uncertainty of about only. By resorting to lower-level accesses to the radio chip such as in (Ramirez et al., 2019), the uncertainty will be further reduced. Thus, the synchronization-free approach can achieve ms or sub-ms timestamping accuracy.

4. Security of Synchronization-Free Timestamping

The synchronization-free approach brings various benefits including implementation simplicity and bandwidth usage saving. Due to the broadcast and long-range nature of LoRaWAN communications, it is also important to investigate the security aspect. In this section, we consider a frame delay attack that maliciously manipulates the propagation time. Thus, the attack will directly affect the synchronization-free timestamping approach that assumes near-zero signal propagation time. In this section, §4.1 defines the threat model; §4.2 presents the implementation of the attack; §4.3 experimentally investigates several important parameter settings to implement the attack; §4.4 discusses a simple attack detection approach and its shortcomings.

4.1. Frame Delay Attack

If an adversary introduces a malicious time delay to the deliveries of the uplink frames from end nodes, the timestamps generated by the gateway will be compromised. We formally define the threat model as follows.

Definition 0 (Frame delay attack).

The end device and gateway are not corrupted by the adversary. However, the adversary may delay the deliveries of the uplink frames from end nodes. The malicious delay for any frame is finite. Moreover, the frame cannot be tampered with because of cryptographic protection.

4.2. Implementation of Frame Delay Attack

This section presents a practical implementation of the frame delay attack via a combination of jamming and replay. §4.2.1 presents the implementation principle; §4.2.2 discusses several practical issues in implementing the attack.

4.2.1. Implementation principle

Figure 1. Implementing frame delay attack by stealthy jamming and replay.

Fig. 1 illustrates the attack implementation. The adversary sets up two malicious devices called eavesdropper and replayer that are close to the end device and the gateway, respectively, to delay the delivery of the uplink frame. The attack consists of three steps. ❶ At the beginning, both the eavesdropper and the replayer listen to the LoRa communication channel between the end device and the gateway. Once the replayer detects an uplink frame transmission from the end device to the gateway, it jams the gateway’s frame reception by transmitting a jamming frame. In §4.3, we will investigate experimentally a stealthy jamming method such that the victim gateway based on an off-the-shelf LoRa radio does not raise any warning message to the application layer. Meanwhile, once the eavesdropper detects an uplink frame transmission from the end device to the gateway, it records the radio waveform of the frame. Note that the replayer may properly control the transmission power of the jamming such that the jamming frame can jam the victim gateway, while not corrupting the radio waveform recorded by the eavesdropper. When the replayer is far away from the eavesdropper, delicate transmission power control of the replayer may be waived because the jamming signal will be weak at the eavesdropper after propagation attenuation. This is experimentally demonstrated in §8. ❷ The eavesdropper sends the recorded radio waveform data to the replayer. ❸ After a time duration of seconds from the onset time of the legitimate frame transmission, the replayer replays the recorded radio waveform received from the eavesdropper. The above jam-and-replay process does not need to decipher the payload of the recorded frame; it simply re-transmits the recorded radio waveform. As the gateway is unaware of the earlier jamming and the integrity of the replayed frame is preserved, the gateway will accept the replayed frame even if it checks the cryptographically protected check sum and frame counter. The above process introduces a delay of seconds to the delivery of the frame.

Although the above descriptions focus on an end device, the attack setup illustrated in Fig. 1 can affect the uplink frames from many end devices close to the eavesdropper, as long as the strength of the signal from an end device at the eavesdropper is much higher than that of the jamming signal from the replayer.

4.2.2. Several practical issues

To increase the stealthiness of the replay attack, the replayer can well control the transmission power of the replay such that only the victim gateway can receive the replayed frame. Although the volume of the recorded radio waveform data is often large, the eavesdropper can transmit the data to the jammer via a separate communication link (e.g., LTE).

As the adversary should delay the uplink frame, how does the adversary know in time the direction of the current transmission? In LoRaWAN, the uplink preamble uses up chirps, whereas the downlink preamble uses down chirps. Thus, the adversary can quickly detect the direction of the current transmission within a chirp time. From our results in §4.3, the jamming should start after several chirps and before tens of chirps of the frame transmission. Thus, a time duration of one chirp for sensing the direction of the transmission will not impede the timeliness of the jamming. And the end device identification can be achieved by using the uplink frame’s source node ID (if not encrypted) or extracting the end device’s frequency trait (see §7.2).

4.3. Attack Experiments

We conduct experiments to investigate the parameter settings for implementing the stealthy jamming attack. In the experiments, we set up two RN2483-based LoRa nodes as the end device and the gateway. We use a third LoRa node as the jammer. The distance between the gateway and the end device is about ; the distance between the jammer and the gateway is about . We primarily investigate the timing of successful stealthy jamming. From our experiments, there are three critical time windows (denoted by , , and ) after the onset time of the legitimate frame transmission from the end device to the gateway (denoted by ). These time windows are illustrated in Fig. 2. If the onset time of the jamming is in , the gateway most likely receives the jamming frame from the jammer only; if it is in , the gateway can receive neither frames and the gateway’s RN2483 raises no alerts; if it is in , the RN2483 reports frame corruption; if it is after , the gateway can receive both frames sequentially. Therefore, the time window is called effective attack window. The jamming attack with an onset time in this window is stealthy to the gateway.

Figure 2. Jamming attack time window.

We measure , , and under various settings for the spreading factor and the payload size of the legitimate frame. Table 1 summarizes the results. From the results for , the jamming should start after the 5th chirp of the legitimate frame transmission. This is because: i) the gateway’s LoRa chip has not locked the legitimate preamble until the 6th chirp and it will re-lock the jamming frame’s preamble due to higher signal strength if the jamming frame starts before the 5th chirp of the legitimate frame; ii) the gateway’s LoRa chip locks the legitimate preamble from the 6th chirp and will simply drop any received radio data without reporting any error if any of the last three chirps (i.e., the 6th, 7th, and 8th chirps) of the preamble and/or the frame header are corrupted. For the latter case of frame header corruption, the hardware cannot determine whether itself is the intended recipient and therefore drops the received data. Thus, the stealthy jamming should start after the 5th chirp of the legitimate frame transmission.

We can also see that increases exponentially with the spreading factor. This is because: i) the total time for transmitting the preamble and frame header increases exponentially with the spreading factor; ii) corruption of the payload after the frame header will lead to an integrity check error and a warning message. The is roughly the time for transmitting the legitimate frame. Thus, if the jamming onset time is after , both the legitimate and the jamming frames can be decoded.

Spreading Chirp Preamble Payload
factor time time (byte)
10 5 28 141
7 1.024 8.2 20 5 38 156
30 6 41 165
40 6 54 178
7 1.024 8.2 6 41 165
8 2.048 16.4 30 10 82 208
9 4.096 32.8 22 156 274
* Unit for chirp time, preamble time, , , is millisecond.
Table 1. Jamming attack time windows for RN2483.

The above experiments show that, there is a time window for the jamming to corrupt the preamble partially and the frame header such that the victim simply drops the received data and raises no alerts. Jamming starting in this window is stealthy.

4.4. Discussion on a Simple Attack Detector

A simple attack detection approach is to perform round-trip timing and then compare the measured round-trip time with a threshold. However, this approach will need a downlink transmission for each uplink transmission, which doubles the communication overhead. LoRaWAN is mainly designed and optimized for uplinks. For instance, a LoRaWAN gateway can receive frames from multiple end devices simultaneously using different spreading factors, whereas it can send a single downlink frame only at a time. This is because Class A specification requires that any downlink transmission must be unicast, in response to a precedent uplink transmission (Sornin et al., 2016). Thus, the round-trip timing approach matches poorly with the uplink-downlink asymmetry characteristic of LoRaWAN. Moreover, as the frame delay attacks will be rare events, continually using downlink acknowledgements to preclude the threat is a low cost-effective solution. Differently, in this paper, we will design advanced signal processing algorithms that run at the gateway to analyze the received radio signals and detect the frame delay attack. Our technique will be a cost-effective solution for the awareness of the attack existence without introducing any communication overhead or any modifications to the hardware/software of the end devices.

5. SoftLoRa Gateway

This section presents the SoftLoRa gateway used to achieve secure data timestamping in LoRaWANs.

5.1. SoftLoRa Hardware

To develop the attack detection capability, we integrate an SDR receiver with a LoRaWAN radio to monitor the LoRa PHY layer. Various cheap (US$25 only (rtl, 2018)) and low-power SDR receivers are widely available now. In this paper, we use RTL-SDR USB dongles based on the RTL2832U chipset (rtl, 2018), which were originally designed to be DVB-T TV tuners. The RTL-SDR supports continuous tuning in the range of MHz, which covers the LoRaWAN bands (i.e., 430, 433, 868, 915 MHz). It can operate at reliably for extended time periods. Thus, the sampling resolution is .

Figure 3. SoftLoRa hardware prototype consisting of Raspberry Pi (host), LoRaWAN shield (LoRa transceiver), RTL-SDR USB dongle (SDR receiver).
Figure 4. Software architecture of SoftLoRa gateway. Bottom part is end device; upper part is the gateway; solid arrows are local data flows; dashed arrows are transmissions.

The research of this paper is conducted based on a SoftLoRa hardware prototype that integrates a Raspberry Pi 3 Model B single-board computer (as the host), a Cooking Hacks LoRaWAN shield (lor, 2018) (as the LoRa transceiver), and an RTL-SDR USB dongle (as the SDR receiver). Fig. 4 shows the prototype. The LoRaWAN shield consists of a Microchip RN2483 chipset, an antenna, and a general-purpose input/output (GPIO) interfacing circuit. RN2483 is based on Semtech SX1276, a major commodity LoRa chip on the current market. Mounted on the host via GPIO pins, the shield can be controlled using a C++ library from Cooking Hacks. An antenna is also integrated with the RTL-SDR to improve signal reception. The RTL-SDR is plugged into a USB port of the gateway’s host computer.

The SDR receiver will be used to capture the radio signal over a time duration of the first two CSS chirps of an uplink frame. The first sampled chirp is used to extract PHY-layer timestamp (cf. §6), whereas the second sampled chirp is used to extract the FB of the transmitter (cf. §7). The microseconds-accurate PHY-layer timestamp is a prerequisite of the FB estimation. As only two chirps’ radio waveform will be analyzed, SoftLoRa will have manageable computation overhead, which can be performed by embedded computing boards such as Raspberry Pi.

An alternative approach is to adopt a full-fledged SDR transceiver (e.g., USRP) to design a highly customized LoRaWAN gateway with PHY signal analysis capability. However, this approach will lose the factory-optimized hardware-speed LoRa demodulation built in the commodity LoRaWAN platforms. Moreover, full-fledged SDR transceivers (e.g., USRP N210) are often 10x more expensive than SoftLoRa. The low-cost, low-power, listen-only RTL-SDR suffices for developing the frame delay attack detector.

5.2. CSS Reception using SDR Receiver

This section models the LoRa signal reception using the SDR receiver. It will be a basis for understanding the challenges of achieving PHY-layer timestamping in §6 and developing accurate FB estimation algorithms in §7.

LoRa adopts CSS modulation. A chirp is a finite-time band-pass signal with time-varying frequency that swaps the whole bandwidth of the communication channel in a linear or non-linear manner. Let and denote the instantaneous amplitude and frequency of the chirp at the time instant . Thus, the chirp, denoted by , is

where is the LoRa transmitter’s phase that is usually unknown.

Figure 5. Analog signal processing in SDR receiver.

Fig. 5 illustrates the essential analog signal processing steps of most SDR receivers to yield the in-phase () and quadrature () components of the received radio signal. The SDR receiver generates two unit-amplitude orthogonal carriers and , where is a specified frequency and is the phase of the two self-generated carriers. The can be set to be the central frequency of the used LoRa channel. The and components, denoted by and , are

(1)
(2)
(3)
(4)

The high-frequency components in Eqs. (2) and (4) are removed by the low-pass filters of the SDR receiver. Thus, the and components after the filtering, denoted by and , are given by Eqs. (1) and (3). They can be rewritten as

The continuous-time and are then sampled by the analog-to-digital converters (ADCs) to yield the and data. For simplicity of exposition, the analysis in this paper is performed in the continuous-time domain.

5.3. SoftLoRa Software Architecture

This section overviews the software architecture of SoftLoRa to achieve secure data timestamping that is resilient to the frame delay attack. It is based on the results in the subsequent sections of this paper.

The upper part of Fig. 4 shows the software architecture of the SoftLoRa. The uplink transmission from the end device is captured by both the gateway’s LoRaWAN transceiver and the SDR receiver. The LoRaWAN transceiver demodulates the received radio signal and passes the frame to the gateway’s computer host. PHY-layer signal processing algorithms are applied on the LoRa signal after down-conversion by the SDR receiver to pick precisely the arrival time of the radio signal (i.e., PHY timestamping), estimate the transmitter’s FB, and detect whether the current frame is a replayed one. The replay detection is by checking whether the estimated FB is consistent with the historical biases associated with the transmitter ID claimed in the current frame. The gateway will be aware of such frame replay attack and drop the replayed frame. Note that SoftLoRa gateway uses the SDR receiver to obtain FBs, rather than to decode the frame.

In this rest of this paper, we will present the signal processing algorithms for PHY signal timestamping in §6, FB estimation and attack detection in §7. Note that microseconds-accurate PHY signal timestamping is a prerequisite of the FB estimation.

6. Signal Timestamping for LoRa

In this section, we present our LoRa signal timestamping approach on SoftLoRa (§6.1) and evaluation results (§6.2). We aim to achieve microseconds accuracy in timestamping the LoRa signal.

6.1. LoRa Signal Timestamping using SDR Receiver

We perform PHY-layer signal timestamping by detecting the onset time of LoRa frame preamble. With preamble onset timestamp, the FB estimation algorithm can select the right segments of and data to work on. In this section, we first model the preamble’s and data received by the SDR receiver and then discuss the preamble onset time detection.

6.1.1. Preamble received by SDR receiver

In LoRaWAN, by default, the preamble of an uplink frame (from end device to gateway) or a downlink frame (from gateway to end device) consists of eight up or down linear chirps, respectively (Sornin et al., 2016). Let and represent the central frequency and bandwidth of the used LoRa channel. In all numerical examples and experiments of this paper, we use a LoRaWAN channel with and (Sornin et al., 2016). The instantaneous frequency of an up chirp increases linearly with time, from the lowest frequency (i.e., ) to the highest frequency (i.e., ) of the channel. It is given by for , where is the spreading factor and is the chirp time. The is an integer within . A larger spreading factor increases the chirp time and thus decreases the data rate when the chirps are used to encode data. However, longer chirp times increase the SNR at the receiver and thus the communication range. By following the analysis in §5.2, the and components of the received up chirp are given by and , where the instantaneous angle is . The analysis for the down chirp with linearly decreasing frequency is similar; we do not elaborate here.

Figure 6. data () and spectrogram of an up chirp.

Fig. 6 shows the data and the spectrogram of an ideal up chirp sampled at . The parameters of the up chirp are , , and . Thus, the chirp time is

. To generate the spectrogram, we apply the short-time fast Fourier transform (FFT) with

-point Kaiser window and 16-point overlap between two neighbor windows. Thus, the spectrogram consists of 20 power spectral densities over the chirp time of .

6.1.2. Preamble onset time detection

Detecting the onset time of the preamble is non-trivial. To understand the challenges, in this section, we discuss two possible methods and their inefficacy. Then, we present two other promising methods.

As the up chirp exhibits a clear time-frequency pattern as shown in Fig. 6, a possible approach to locating the first up chirp of the preamble is to analyze the spectrogram of the received and data. However, the spectrogram inevitably has reduced time resolution. For instance, the time resolution of the spectrogram in Fig. 6 is , which impedes high-resolution PHY-layer timestamping.

Figure 7. data of an up chirp with (the red curve) or (the blue curve).
Figure 8. Actual data of an up chirp. The time shift of the dip center is caused by FB.

Matched filter is a widely adopted symbol detection technique for constant carrier frequency modulation schemes, e.g., ASK and PSK. As a coherent detection technique, the matched filter requires that the receiver is phase-locked to the transmitter (i.e., ) to achieve the best symbol detection accuracy. However, as LoRa adopts time-varying frequency, it is difficult for the SDR receiver to estimate the transmitter’s phase . In fact, low-end SDR receivers such as the RTL-SDR used in this work do not provide phase-lock capability. As a result, the phase difference , which is a critical factor affecting the shape of and , will be random. Fig. 8 shows the ideal traces of the up chirp when the is and . The waveform shapes are different. Thus, it is impossible to define a template shape for the matched filter to work. Moreover, as analyzed in §7, the FB of the LoRa transmitter will significantly alter the shapes of the and signals. Fig. 8 shows the actual trace captured by the SDR receiver. The dip center shift is caused by FB. Thus, due to the random and the LoRa transceiver’s FB, the matched filter is not promising.

(a) Envelope detector
(b) AIC detector
Figure 9. Preamble onset time detection results.

To investigate the signal timestamping of LoRa, we consider two time-domain signal processing techniques:

Envelope detector: First, we apply the Hilbert transform to extract the amplitude envelope of the or signal. Then, the sample with the largest ratio between its envelope amplitude and the previous sample’s envelope amplitude is yielded as the preamble onset. Fig. 9(a) shows the extracted amplitude envelope and the ratio curve over time. We can see that the maximum ratio well indicates the onset time of the preamble.

AIC detector: The autoregressive Akaike Information Criterion (AIC) algorithm (Sleeman and Van Eck, 1999) has been widely adopted to estimate the arrival time of seismic wave with an accuracy of a single sampling point. As the and

signals are similar to the seismic waves, the AIC is a promising solution for our problem. The algorithm works as follows. For each point of the signal as an onset time candidate, two autoregressive models are constructed for the signal segments before and after the onset time candidate. The candidate that gives the largest dissimilarity between the two autoregressive models is yielded as the final onset time estimate. The vertical line in Fig. 

9(b) represents the onset time detected by the AIC detector.

Note that as both the envelope detector and the AIC detector formulate the onset time detection as optimization problems, they do not need any detection threshold.

6.2. Signal Timestamping Accuracy Evaluation

ENV I error 5.4 4.5 4.8 5.2 1.9 5.2 4.6 3.5 6.3 5.2
Q error 6.3 6.7 5.6 7.3 7.1 4.6 5.4 5.2 5.2 9.8
AIC I error 1.0 1.3 1.5 0.8 1.7 1.7 1.5 1.3 1.3 1.9
Q error 1.5 0.6 1.5 1.5 1.3 0.6 1.0 1.5 1.0 1.5
Table 2. Error upper bound (s) by envelope (ENV) detector and AIC.

We conduct a set of experiments to evaluate the accuracy of our PHY-layer signal timestamping. The accuracy of signal timestamping is restricted by the sampling rate of the SDR. When the real onset time is between two consequent samples, the real onset time is unknown while the range can be confirmed. Thus, we use the upper bound of the signal timestamping error to evaluate our approach. In §8, we will also measure this error metric in a multistory building and a campus, where the signal propagation will be affected by the noise and the attenuation with distance.

Tables 2 shows the error upper bound measured for the envelope detector and the AIC detector when operating on the and data in ten experiments, respectively. We can see that the AIC detector achieves higher accuracy. In particular, the timing errors of the AIC detector are less than .

Figure 10. AIC timestamping error vs. received SNR.

Then, we evaluate the impact of random noises on AIC’s signal timestamping accuracy. We artificially add zero-mean Gaussian noises to the collected high-SNR and traces. Then, we apply the AIC detector on the noise-added traces to detect the LoRa signal onset time. The SNR in dB is defined as . Fig. 10 shows the results. From our measurements in a multistory building (cf. §8), the received SNR ranges from to . From Fig. 10, the average timestamping error is expected to be within . This will be confirmed by real experiments in §8. When the SNR is , which is the lower limit for reliable demodulation (sx1, 2018), from Fig. 10, the average error will be within .

7. Frequency bias-based frame delay attack detection

Internal oscillators for generating carriers generally have frequency biases (FBs) of one to hundreds of ppm, due to manufacturing imperfection. This section develops algorithms for estimating LoRa transmitters’ FBs. We also investigate whether such FBs can be used to detect the replay attacks.

7.1. FB Estimation

This section describes algorithms for estimating the transmitter’s FB based on an up chirp in the preamble. The method for a down chirp is similar. First, we analyze the impact of the transmitter’s and SDR receiver’s FBs (denoted by and ) on the and traces. The up chirp’s instantaneous frequency accounting for is

The two local unit-amplitude orthogonal carriers generated by the SDR receiver are and . Following the analysis in §5.2, the and components of the received up chirp can be derived as and , where the instantaneous angle is given by

(5)

The difference between the transmitter’s and receiver’s FBs, i.e., , affects the and waveforms. Fig. 11 shows the numerical results of traces when and . The non-zero shifts the axis of symmetry of the trace. We observed this in the actual data shown in Fig. 8.

(a)
(b)
Figure 11. . All x-axes are time in ms; =125kHz; =7.

For a certain SDR receiver, the FB estimation problem is to estimate from the captured and traces. We do not need to estimate , because for a certain SDR receiver with a nearly fixed , a change in indicates a change in

and a replay attack. In what follows, we describe two approaches based on linear regression and least squares formulations. The least squares approach keeps robust when the SNR is very low but has higher computation overhead.

7.1.1. Linear regression approach

From Eq. (5), the FB appears in the linear item of only. Thus, the is a linear function of , i.e., . The slope of , i.e., , can be estimated by linear regression based on the data pairs , where , , and rectifies the multi-valued inverse tangent function to an unlimited value domain. The rectification is as follows. The is initialized to be when . As increases, if jumps from to , decreases by one; if jumps from to , increases by one. Note that the traces and where are the segments of the captured and signals starting from the preamble onset time detected by the AIC algorithm and lasting for a chirp time duration of seconds.

(a) and
(b)
(c) with rectification
(d)
Figure 12. Estimating FB from real and traces. All x-axes are time in ms from the chirp onset time.

Fig. 12 shows the intermediate results of the FB extraction. Fig. 12(a) shows real and traces of an up chirp emitted by an RN2483-based LoRa transmitter and captured by the SDR receiver. Fig. 12(b) shows the . Fig. 12(c) shows the obtained by rectifying the result in Fig. 12(b) with . Fig. 12(d) shows . We can see that it is indeed a linear function of time , which conforms to our analysis. By applying linear regression to the result in Fig. 12(d), the FB (i.e., the slope of the fitted line divided by ) is estimated as . Note that the nominal central frequency is . The FB is merely 26 ppm of the central frequency.

Figure 13. FBs estimated from the original LoRa signals from 16 nodes and those replayed by a USRP-based replayer. The error bar shows mean, minimum, and maximum of FBs in 20 frame transmissions.

We use an SDR receiver to estimate the FBs of 16 RN2483-based LoRa transmitters. In each test for a LoRa transmitter, the distance between the transmitter and the SDR receiver is about . The error bars labeled “original” in Fig. 13 show the results. We can see that the FBs for a certain node are stable and the nodes generally have different FBs. The absolute FBs are from to , which are about to of the nominal central frequency of .

Figure 14. Frequency estimation error of the least squares-based approach under different SNRs of the received signal.

Some nodes have similar FBs, e.g., Node 3, 8, and 14. Recall our discussion in §4.2.1 that the adversary may wish to identify the nodes based on FBs to selectively attack a certain node. To address the issue of similar FBs, the adversary may jointly use the FBs and the received signal strengths that are affected by the transmitters’ geographic locations to fingerprint the transmitters. Differently, the detection of the replay attack is based on the fact that the replayed transmission will have a different FB. In other words, the attack detection does not require distinct FBs among different transmitters.

As the linear regression approach has a closed-form formula to compute , it has a complexity of in the search of the solution. However, the inverse tangent rectification is susceptible to low received SNRs. In next section, we present the least squares approach that still performs well when the received SNR is below the limit for reliable demodulation, but with a higher complexity in searching for the solution.

7.1.2. Least squares approach

The LoRa signals can be very weak after long-distance propagation or barrier penetration. The LoRa’s demodulation is designed to address low SNRs. For Semtech SX1276, the minimum SNRs required for reliable demodulation with spreading factors from 7 to 12 are from to (sx1, 2018). We aim at extracting FB at comparably lower SNRs. We solve a least squares problem:

where and are the received and traces; is given by Eq. (5); and are the noiseless and templates. The above formulation requires that the and templates have an identical and constant amplitude . From our measurements, this requirement is met when the second chirp of the preamble is used. The can be estimated as follows. Let and , where and are zero-mean random noises in the and traces, respectively. Thus, the average power of the LoRa signal can be derived as , where is the average noise power that can be measured when there is no LoRa signal. Thus, can be estimated using the average powers of the noisy LoRa signal and the pure noise.

We use a scipy implementation of the differential evolution algorithm (Storn and Price, 1997) to solve the least squares problem. On the SoftLoRa platform, it takes 0.69 seconds to solve it. For evaluation, we artificially add noise traces to the high-SNR and traces used in §7.1. By controlling the magnitudes of the added noise traces, we can achieve a certain SNR in dB defined as . We use two types of noise traces: randomly generated zero-mean Gaussian noise traces and real noise traces captured using an SDR receiver in a multistory building. The amplitudes of the real noise traces are scaled to achieve different SNRs. From Fig. 14, we can see that the FB estimation errors are below (i.e., ), when the SNR is down to for both types of noises.

7.2. Replay Attack Detection

The malicious replayer also has an FB. We use a USRP N210 SDR transceiver to build a replayer. The error bars labeled “replayed” in Fig. 13 show the FBs estimated from the LoRa signals received by the SoftLoRa’s SDR receiver when the USRP replays the radio waveform captured by itself in the experiments presented in §7.1. Compared with the results labeled “original”, the FBs of the replayed transmissions are consistently lower. This is because the USRP has a negative FB. The average additional FBs introduced by the replayer range from to , i.e., to of the channel’s central frequency. Thus, with the FB estimation accuracy of achieved under low SNRs in §7.1.2, the additional FBs caused by the replay attack can be detected.

Based on the above observation, we describe an approach to reliably detect the replay attacks. We assume that a SoftLoRa gateway has a database of the FBs of the nodes with which it communicates. This database can be built offline or at run time using its SDR receiver in the absence of attacks. To address the neighbors’ time-varying radio frequency skews due to run-time conditions like temperature, the SoftLoRa gateway can continuously update the database entries based on the FBs estimated from recent frames. To decide whether the current received frame is a replayed frame, the SoftLoRa gateway checks whether the FB of the current received frame is within the FB range of the claimed source node in the database. This detection approach is applied after the SoftLoRa gateway decodes the frame to obtain the claimed source node ID. The FB estimated from a frame that is detected to be a replayed one should not be used to update the database. Through the FB monitoring, the SoftLoRa gateway can detect the replay attack and will not use the replayed frame to do data timestamping. Thus, the data timestamps will not be spoofed by the frame delay attack.

We discuss two notes about the detection mechanism. First, to bypass the above detection mechanism, the attacker will need SDRs with FBs within (i.e., the SoftLoRa’s FB estimation accuracy). However, as RF devices typically have FBs of one to tens of ppm, it is difficult to bypass the detection mechanism. Second, the detection does not require uniqueness of the FBs across different LoRa transceivers, because it is based on changes of FB.

8. Performance Evaluation

8.1. Experiments in a Multistory Building

Figure 15. SNR survey in the building (lateral view) and signal timing accuracy. The building has three sections (A, B, C) and two junctions (J). The C3 positions on the 1st and 2nd floors are not accessible. The triangle indicates the position of the fixed node. The numbers in the cells are the measured timing error upper bounds in s when the mobile node is at the corresponding locations.

We conduct experiments in a concrete building with six floors. Along its long dimension of 190 meters, the building has three sections and two section junctions. Fig. 15 illustrates a lateral view of the building. We survey the SNR inside the building. We deploy a fixed LoRaWAN transmitter in Section A on the 3rd floor, as illustrated by the triangle in Fig. 15. Then, we carry a mobile SoftLoRa receiver to different positions inside the building to measure the SNR. At each position, we first profile the noise power and then measure the total power when the fixed node transmits. We use the method described in §7.1.2 to compute SNR. In each section, we measure three positions. The heat map in Fig. 15 shows the SNR measurements. We can see that the SNR decays with the distance between the two nodes. The SNRs are from to . Then, we conduct the following experiments. Default settings used in our experiments are: spreading factor ; ; .

8.1.1. Full implementation of frame delay attack

We deploy a commodity LoRaWAN end device and a SoftLoRa gateway in Section A1 of the 3rd floor and Section C3 of the 6th floor, respectively. The LoRa signals are significantly attenuated after passing through multiple building floors. If the end device adopts a spreading factor of 7, it cannot communicate with the gateway due to the signal attenuation. A minimum spreading factor of 8 is needed for reliable data communications between them. In rural environments, this spreading factor can be used to achieve communication ranges of three to four kilometers (Mikhaylov et al., 2016). Thus, our multistory building environment creates realistic challenges similar to those caused by long geographic distances. Following the attack scheme in Fig. 1, we deploy two USRP N210 stations as the eavesdropper and the replayer, next to the end device and the gateway, respectively. We fully implement the attack steps described in §4.2.1. We set the transmission power of the master to the maximum level, i.e., 15. The transmission power of the jamming signal is . After crossing multiple building levels, the jamming signal arriving at the eavesdropper is weak. As a result, when the replayer replays the and data captured by the eavesdropper that contains the weak jamming signal, the gateway can successfully decode the frame. Moreover, if the replayer’s USRP adopts a transmission power no greater than , the replayed frame cannot be detected by the gateway, making the replay attack stealthy to the gateway. This experiment shows the credibility of the frame delay attack in a building.

8.1.2. Signal timestamping accuracy

We measure the signal timestamping error metric defined in §6.2 when the mobile node is at different locations in the building. Note that this signal timestamping is a prerequisite of the subsequent FB estimation. The numbers shown in the cells of Fig. 15 are the measured timestamping error metric in s when the mobile node is at the corresponding locations. SoftLoRa achieves sub-10s signal timestamping accuracy in a concrete building. These results are consistent with our earlier results in Fig. 10 with SNRs from to (i.e., the SNR range in this building).

Figure 16. Estimated FB vs. transmission power of the end device. Each box plot shows the minimum, maximum, 25% and 75% percentiles.

8.1.3. Impact of transmission power on FB estimation

Fig. 16 shows the estimated FBs versus the end device’s transmission power under different settings. The bottom row of black box plots are the FBs estimated by the eavesdropper when the end device transmits the uplink frame with different transmission powers. The middle row of red box plots are the FBs estimated by the SoftLoRa gateway in the absence of the jamming and replay attacks. Thus, the FBs estimated by the eavesdropper and the SoftLoRa gateway are different. This is because that as analyzed in §7.1, the estimated FB contains the transmitter’s and receiver’s FBs and . Note that the eavesdropper and the SoftLoRa gateway in general have different FBs. From Fig. 16, the end device’s transmission power has little impact on the FB estimation.

8.1.4. Additional FB introduced by replayer

In Fig. 16, the top row of blue box plots are the FBs estimated by the SoftLoRa gateway when the replayer replays the radio waveform recorded by the eavesdropper. We set the gain of the replayer’s USRP to be . When the end device adopts a higher transmission power, the replayed signal also has higher power. By comparing the middle and the top rows, we can see that the replay attack introduces an additional FB of about , which is of the LoRa channel’s central frequency. Therefore, the FB monitoring can easily detect the replay attack. Compared with the results in Fig. 13 showing additional FBs of to , the FBs in this set of experiments are higher. This is because that here we use two different USRPs as the eavesdropper and replayer; their FBs are superimposed.

8.2. Long-Distance Experiments in a Campus

We deploy a LoRaWAN end device and a SoftLoRa gateway in a campus, which are separated by an Euclidean distance of . The signal’s one-way propagation time is . The end device and the SoftLoRa gateway are deployed on the roof top of a building and an open stair case of another building, respectively. Fig. 17(c) shows the two sites (Site A and B) for deploying the end device and the SoftLoRa gateway at the campus. Site A is at the roof top of a building, whereas Site B is in an open stair case of another building. Fig. 17 shows the pictures taken at the two sites. The circled construct in a figure is the building where the other site is located in. We conducted four tests to evaluate the signal timestamping error. It rained heavily during the tests. The measured error upper bounds during the four tests are , , , . We can see that SoftLoRa gateway achieves microseconds signal timestamping accuracy over a distance of one kilometer. This will ensure accurate FB estimation.

(a) View Site B from Site A
(b) View Site A from Site B
(c) The deployment sites
Figure 17. Pictures taken at the two sites.

9. Conclusion

This paper considers the security of a synchronization-free data timestamping approach for LoRaWANs. Specifically, the timestamping of data is performed by the LoRaWAN gateway based on the frame arrival time. The low communication overhead of this approach makes it suitable for the bandwidth-limited LoRaWANs for collecting low-rate data from geographically distributed end devices. However, we show that this approach is susceptible to a basic threat of frame delay attack that can be implemented by a combination of stealthy jamming and delayed replay. As the attack does not need to break the cryptographic protection of the frame, existing security measures prescribed by LoRaWAN cannot counteract this threat. To address this attack, we propose a gateway design called SoftLoRa that integrates a low-power SDR receiver with a LoRaWAN gateway. We design efficient time-domain signal processing algorithms based on the CSS modulation method to estimate the frequency biases of the end devices. Our algorithms achieve a resolution of 0.14 ppm of the carrier frequency, which is sufficient to uncover the additional frequency biases introduced by the replay step of the frame delay attack. In summary, with SoftLoRa gateway, we present a cost-effective defense approach for the low-overhead synchronization-free data timestamping approach against the frame delay attack.

Acknowledgements.
This research was funded by a Start-up Grant at Nanyang Technological University (NTU). We acknowledge Zhenyu Yan for assistance in conducting the long-distance experiments in the NTU campus.

References

  • (1)
  • lor (2018) 2018. LoRaWAN Shield for Raspberry Pi. https://www.cooking-hacks.com/lorawan-shield-for-raspberry-pi-868-mhz-xbee-socket.
  • rtl (2018) 2018. RTL-SDR. https://www.rtl-sdr.com/.
  • sx1 (2018) 2018. SX1276 Datasheet. http://modtronix.com/prod/components/wireless/sx1276.pdf.
  • Aras et al. (2017a) E. Aras, G. S. Ramachandran, P. Lawrence, and D. Hughes. 2017a. Exploring the Security Vulnerabilities of LoRa. In The 3rd IEEE International Conference on Cybernetics (CYBCONF). 1–6.
  • Aras et al. (2017b) Emekcan Aras, Nicolas Small, Gowri Sankar Ramachandran, Stephane Delbruel, Wouter Joosen, and Danny Hughes. 2017b. Selective Jamming of LoRaWAN using Commodity Hardware. In The 14th EAI International Conference on Mobile and Ubiquitous Systems: Computing, Networking and Services (MobiQuitous).
  • Byers ([n. d.]) Charles Byers. [n. d.]. Fog Computing for Industrial Automation. https://www.controleng.com/articles/fog-computing-for-industrial-automation/.
  • Dongare et al. (2018) Adwait Dongare, Revathy Narayanan, Akshay Gadre, Anh Luong, Artur Balanuta, Swarun Kumar, Bob Iannucci, and Anthony Rowe. 2018. Charm: Exploiting Geographical Diversity Through Coherent Combing in Low-Power Wide-Area Networks. In The 17th ACM/IEEE International Conference on Information Processing in Sensor Networks (IPSN).
  • Eletreby et al. (2017) Rashad Eletreby, Diana Zhang, Swarun Kumar, and Osman Yağan. 2017. Empowering Low-Power Wide Area Networks in Urban Settings. In ACM SIGCOMM.
  • Gu et al. (2018) Chaojie Gu, Rui Tan, Xin Lou, and Dusit Niyato. 2018. One-Hop Out-of-Band Control Planes for Low-Power Multi-Hop Wireless Networks. In The 37th Annual IEEE International Conference on Computer Communications (INFOCOM). 1187–1195.
  • Hao et al. (2011) Tian Hao, Ruogu Zhou, Guoliang Xing, and Matt Mutka. 2011. WizSync: Exploiting Wi-Fi infrastructure for clock synchronization in wireless sensor networks. In The 32nd IEEE Real-Time Systems Symposium (RTSS). IEEE, Vienna, Austria, 1379–1392.
  • Hessar et al. (2019) Mehrdad Hessar, Ali Najafi, and Shyamnath Gollakota. 2019. NetScatter: Enabling Large-Scale Backscatter Networks. In The 16th USENIX Symposium on Networked Systems Design and Implementation (NSDI).
  • Institute ([n. d.]) The European Telecommunications Standards Institute. [n. d.]. Technical characteristics for Low Power Wide Area Networks Chirp Spread Spectrum (LPWAN-CSS) operating in the UHF spectrum below 1 GHz. https://www.etsi.org/docdeliver/etsi_tr/103500_103599/103526/01.01.01_60/tr_103526v010101p.docx.
  • Marcelis et al. (2017) Paul J Marcelis, Vijay S Rao, and R Venkatesha Prasad. 2017. DaRe: Data recovery through application layer coding for LoRaWAN. In The 2nd ACM/IEEE International Conference on Internet-of-Things (IoTDI).
  • Mikhaylov et al. (2016) K. Mikhaylov, Juha Petaejaejaervi, and T. Haenninen. 2016. Analysis of Capacity and Scalability of the LoRa Low Power Wide Area Network Technology. In The 22th European Wireless Conference. 1–6.
  • Mizrahi (2014) T. Mizrahi. 2014. Security Requirements of Time Protocols in Packet Switched Networks. https://tools.ietf.org/html/rfc7384.
  • Peng et al. (2018) Yao Peng, Longfei Shangguan, Yue Hu, Yujie Qian, Xianshang Lin, Xiaojiang Chen, Dingyi Fang, and Kyle Jamieson. 2018. PLoRa: a passive long-range data network from ambient LoRa transmissions. In ACM SIGCOMM. ACM, 147–160.
  • Petrić et al. (2016) T. Petrić, M. Goessens, L. Nuaymi, L. Toutain, and A. Pelov. 2016. Measurements, performance and analysis of LoRa FABIAN, a real-world implementation of LPWAN. In IEEE 27th Annual International Symposium on Personal, Indoor, and Mobile Radio Communications (PIMRC). 1–7.
  • Ramirez et al. (2019) Ceferino Gabriel Ramirez, Anton Sergeyev, Assya Dyussenova, and Bob Iannucci. 2019. LongShoT: long-range synchronization of time. In Proceedings of the 18th International Conference on Information Processing in Sensor Networks. ACM, 289–300.
  • Raza et al. (2017) Usman Raza, Parag Kulkarni, and Mahesh Sooriyabandara. 2017. Low power wide area networks: An overview. IEEE Communications Surveys & Tutorials 19, 2 (2017), 855–873.
  • Robyns et al. (2017) Pieter Robyns, Eduard Marin, Wim Lamotte, Peter Quax, Dave Singelée, and Bart Preneel. 2017. Physical-layer fingerprinting of LoRa devices using supervised and zero-shot learning. In The 10th ACM Conference on Security and Privacy in Wireless and Mobile Networks (WiSec). ACM.
  • Sleeman and Van Eck (1999) Reinoud Sleeman and Torild Van Eck. 1999. Robust automatic P-phase picking: an on-line implementation in the analysis of broadband seismogram recordings. Physics of the earth and planetary interiors 113, 1-4 (1999), 265–275.
  • Sornin et al. (2016) N. Sornin, M. Luis, T. Eirich, T. Kramp, and O. Hersent. 2016. LoRaWAN™Specification (V1.0.2).
  • Storn and Price (1997) Rainer Storn and Kenneth Price. 1997.

    Differential evolution–a simple and efficient heuristic for global optimization over continuous spaces.

    Journal of global optimization 11, 4 (1997), 341–359.
  • Talla et al. (2017) Vamsi Talla, Mehrdad Hessar, Bryce Kellogg, Ali Najafi, Joshua R Smith, and Shyamnath Gollakota. 2017. LoRa Backscatter: Enabling the vision of ubiquitous connectivity. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies (IMWUT) 1, 3 (2017), 105.
  • Tay (2018) Eileen Hui Lian Tay. 2018. Cybersecurity of long-range wide-area networks.
  • Toldov et al. (2016) Viktor Toldov, J.P. Meijers, Roman Igual-Perez, Riaan Wolhuter, Nathalie Mitton, and Laurent Clavier. 2016. Performance evaluation of LoRa radio solution for PREDNET wildlife animal tracking project. In 1st Annual LPWAN Conference.
  • Varshney et al. (2017) Ambuj Varshney, Oliver Harms, Carlos Pérez-Penichet, Christian Rohner, Frederik Hermans, and Thiemo Voigt. 2017. LoRea: A Backscatter architecture that achieves a long communication range. In The 15th ACM Conference on Embedded Networked Sensor Systems (SenSys). ACM, 18.
  • Zhu (2017) Jiahui Zhu. 2017. Cybersecurity of long-range wide-area networks.