PaperlayerView exact revision on arXiv

Source-linked AI summary

Injected and Delivered: Fabricating Implicit Control over Actuation Systems by Spoofing Inertial Sensors

Yazhou Tu, Zhiqiang Lin, Insup Lee, Xiali Hei

arXiv:1806.07558v2cs.CR

TL;DR

Embedded inertial sensors provide feedback for timely control, but prior acoustic attacks rarely demonstrated effective control in real systems under realistic conditions. This paper develops non-invasive out-of-band spoofing methods and tests them across devices. The attacks achieve implicit control on 17 of 25 tested devices, and the paper reports that user involvement and moving targets can constrain attack effects.

  • Problem

    Prior attacks rarely demonstrated effective control over embedded inertial sensors while accounting for realistic factors such as inaccessible sampling information and sample-rate drift.

    Source: §1 · p. 1

  • Method

    The authors develop non-invasive acoustic spoofing attacks that manipulate embedded sensor outputs and derived inertial information, using Side-Swing and Switching methods.

    Source: §1 · p. 2

  • Results

    17 of 25 tested devices were susceptible to implicit control, while 23 could be affected by acoustic signals.

    Source: §1 · p. 2

  • Takeaways & Limitations

    A case study shows that signals with relatively low frequencies through a universal sensing channel could manipulate the oscillating digitized signal.

    Source: §1 · p. 2, §8.4 · p. 15

    Moving targets may require tracking and beam alignment, while user involvement can counter or disrupt attack effects on closed-loop scooters.

    Source: §8.3 · pp. 14–15

Abstract

from arXiv · show

Inertial sensors provide crucial feedback for control systems to determine motional status and make timely, automated decisions. Prior efforts tried to control the output of inertial sensors with acoustic signals. However, their approaches did not consider sample rate drifts in analog-to-digital converters as well as many other realistic factors. As a result, few attacks demonstrated effective control over inertial sensors embedded in real systems. This work studies the out-of-band signal injection methods to deliver adversarial control to embedded MEMS inertial sensors and evaluates consequent vulnerabilities exposed in control systems relying on them. Acoustic signals injected into inertial sensors are out-of-band analog signals. Consequently, slight sample rate drifts could be amplified and cause deviations in the frequency of digital signals. Such deviations result in fluctuating sensor output; nevertheless, we characterize two methods to control the output: digital amplitude adjusting and phase pacing. Based on our analysis, we devise non-invasive attacks to manipulate the sensor output as well as the derived inertial information to deceive control systems. We test 25 devices equipped with MEMS inertial sensors and find that 17 of them could be implicitly controlled by our attacks. Furthermore, we investigate the generalizability of our methods and show the possibility to manipulate the digital output through signals with relatively low frequencies in the sensing channel.

1 Introduction

Inertial sensors let control systems respond promptly to motion, but embedded-sensor attacks have rarely demonstrated effective control under realistic conditions. This work develops non-invasive acoustic spoofing methods and evaluates their effects on real systems.

1 Introduction

  • Prior work left open how far attackers could exploit embedded inertial sensors, especially without access to sampling or sensor data and despite sample-rate drift.The paper also targets systems that use measurements in both directions, unlike prior control of a vehicle in only one direction.Source: §1 · p. 1
  • Out-of-band acoustic signals are undersampled, so small ADC sample-rate drifts can shift digital-signal frequency; amplitude adjustment and phase pacing offer ways to control the oscillating output.Switching the analog-signal frequency can induce a phase offset in the digitized signal.Source: §1 · p. 2
  • 17 of 25 tested devices were susceptible to implicit control, and 23 could be affected by acoustic signals.Demonstrations included self-balancing transporters, virtual reality, and navigation systems.Source: §1 · p. 2
  • The authors devise non-invasive attacks that manipulate embedded sensor outputs and derived inertial information to deceive control systems.Their listed contributions include Side-Swing and Switching attacks, and adversarial control demonstrations across several system types.Source: §1 · p. 2
  • A case study indicates that relatively low-frequency signals sent through a universal sensing channel can manipulate the oscillating digitized signal.This is presented as an investigation of the methods’ generalizability.Source: §1 · p. 2

2 Inertial Sensors in Control Systems

MEMS inertial sensors translate mechanical motion into electrical signals that control systems use for automated decisions. The section describes how acoustic resonance can enter this signal path and distinguishes closed-loop from open-loop control.

2 Inertial Sensors in Control Systems

  • MEMS accelerometers use a mass-spring structure to detect linear acceleration, while gyroscopes use a driven vibrating structure to sense Coriolis acceleration.For gyroscopes, angular velocity is derived from Coriolis acceleration and the structure’s velocity.Source: §2 · p. 2
  • Acoustic signals near a sensor’s natural frequency can drive its sensing mass into resonance, producing electrical signals that the ADC digitizes before they reach the control system.The sensing mass’s displacement is measured by capacitive electrodes.Source: §2 · p. 2
  • Attackers are constrained to sinusoidal analog signals from mechanical resonance and cannot directly control the digital signal, making embedded signal control difficult.They must induce specific digital signals indirectly by controlling the analog input.Source: §2 · p. 2
  • Inertial sensors support diverse control applications, including transporters, robots, navigation, stabilization, and industrial automation.The paper emphasizes their role in timely feedback for critical decisions.Source: §2 · p. 3
  • Closed-loop systems compare current status with a goal and actuate to reduce the difference, whereas open-loop systems follow inertial information to make actuation decisions.Source: §2 · p. 3

3 Threat Model

The threat model aims to spoof embedded inertial sensors and induce digital signals that trigger system actuations. It assumes non-invasive acoustic access, with attackers unable to alter hardware or directly access the sampling process or sensor output.

3 Threat Model

  • The attacker’s objective is to spoof an embedded inertial sensor and induce digital signals that trigger actuations.Source: §3 · p. 3
  • The attack is non-invasive: attackers cannot alter hardware or directly access or modify sampling or sensor output, but may analyze an identical device beforehand.Source: §3 · p. 3
  • Gyroscope attacks often use inaudible ultrasound, while accelerometer resonant frequencies are usually within human hearing.Gyro resonances in several tested devices were 19–20 kHz, above the audible range of most adults.Source: §3 · p. 3
  • The model assumes consumer-grade sound equipment and a possible attack distance of several meters, with attackers able to optimize the sound source.More capable attackers may use professional acoustic devices or customized amplification.Source: §3 · p. 3

4 Modeling and Analysis

This section models how out-of-band acoustic signals are digitized and develops two ways to shape the resulting sensor output: adjusting sample amplitudes and pacing phase by changing frequency.

4.1 Digitization of Out-of-band Signals

  • A 0.01-Hz sample-rate drift yields a −1-Hz sampled-signal frequency deviation when F=20,000 Hz and FS=200 Hz.More generally, when FS<2F, a drift ΔFS is amplified to a −n·ΔFS frequency deviation; Figure 2 shows a nonconstant output under fixed-frequency injection.Source: §4.1 · p. 4, Figure 2 · p. 4

4.2 Digital Amplitude Adjusting

  • Undersampling lets different sample amplitudes be adjusted to reshape the digital signal rather than preserve its oscillating pattern.When F is considerably larger than FS/2, adjacent sample amplitudes become less dependent, enabling individual adjustment and fabrication of specific waveforms.Source: §4.2 · p. 4, Figure 3 · p. 4

4.3 Phase Pacing

  • Phase pacing switches the out-of-band frequency to induce a phase offset; opposite-sign aliased frequencies reverse the digital signal’s moving direction.Figure 4 illustrates the direction reversal when the phase derivative changes sign.Source: §4.3 · pp. 4–5, Figure 4 · p. 5

4.4 Out-of-band Signal Injection Model

  • The injection model identifies sample amplitudes A[i] and aliased frequency ε as controllable parameters for shaping digitized output.Adjusting A[i] changes sample values proportionally, while changing the analog frequency alters ε; flipping ε’s sign inverts the digital signal’s moving direction.Source: §4.4 · p. 5

5 Attack Methods

The attack methods manipulate embedded inertial readings and derived heading angles to influence control systems. Side-Swing shapes signal amplitude, while Switching paces signal phase to sustain directional heading accumulation.

5.1 Side-Swing Attack

  • Side-Swing solves oscillation’s zero net heading per cycle by amplifying samples in the target direction and suppressing those in the opposite direction.Its heading grows during only half of each signal period, which can limit the angle accumulated over a given time.Source: §5.1 · p. 6, Figure 5 · p. 6
  • 17.6 rads (1008°) accumulated in about 25 seconds under Side-Swing, with an average-to-peak angular-speed ratio of 0.15.The iPhone 5 was stationary during the test, and the attack induced output mainly in the target direction.Source: §5.1 · p. 6, Figure 6 · p. 6

5.2 Switching Attack

  • Switching repeatedly alternates acoustic frequencies to pace digital-signal phase, keeping the induced angular speed directed so heading continues to grow.The target can be positive or negative; in practice, the induced signal’s |ε| should stay below 0.5 or 1, depending on the attacker’s reaction speed.Source: §5.2 · pp. 6–7
  • 6.5 rads (372.4°) accumulated in about 25 seconds under Switching, with an average-to-peak angular-speed ratio of 0.58 versus 0.15 for Side-Swing.The iPhone 7 was stationary during the test; the Switching attack used two acoustic frequencies and could produce a larger heading angle.Source: §5.2 · p. 7, Figure 7 · p. 7

6 Evaluations

Across 25 tested devices, 17 were susceptible to implicit control; the evaluations span attack preparation and both closed- and open-loop systems. The results also show differing vulnerabilities across device types and sensors.

6 Evaluations

  • 17 of 25 tested devices were susceptible to implicit control; 2 others had limited control, 4 were vulnerable to denial of service, and 2 were unaffected.The evaluation covered a range of real systems using MEMS inertial sensors.Source: §6 · p. 7, Table 1 · p. 10, Table 2 · p. 12

6.1 Attack Overview

  • Attackers can infer induced sensor direction and magnitude from system actuations, then tune acoustic signals using observed feedback.Profiling identifies resonant frequencies, and attacks adjust amplitude or switch frequencies according to the induced actuation.Source: §6.1 · pp. 7–8

6.2 Experimental Setup

  • The sound source combines multiple tweeter speakers to avoid frequency-band weaknesses and select the speaker with the highest measured SPL.The setup measures speakers from 1.5 to 31.5 kHz and uses a directivity horn during attacks.Source: §6.2 · p. 9

6.3 Experiments on Closed-loop Systems

  • Two scooters allowed control of motor direction and speed; tested robots and other transporters showed weaker or different attack outcomes.The Segway One S1 tilted only slightly, the Minilite tended to lose control, Mitu sped up under spoofed rotations, and MiP’s accelerometer was vulnerable to denial of service.Source: §6.3 · pp. 9–10
  • Both tested stabilizers were vulnerable to control, while the Liftware handle suffered denial of service and the Gyenno spoon was unaffected.The Osmo was mainly affected on X-axis and the Wenpod on Y-axis; gravity-based calibration increasingly opposed fabricated heading angles.Source: §6.3 · p. 10

6.4 Experiments on Open-loop Systems

  • The ES120 screwdriver’s motor direction and speed were controllable, while the B&D required panel removal and the Dewalt was unaffected.The Hololens was affected only slightly, while Oculus Rift view and Oculus Touch object orientation could be manipulated on X-axis.Source: §6.4 · p. 11
  • All six tested smartphones and both 3D mice were vulnerable to adversarial control through ultrasonic signals.Attacks manipulated VR views and aiming, or pointed remote-system cursors at different targets.Source: §6.4 · p. 11
  • Acoustic denial-of-service attacks repeatedly woke a Pro32 soldering iron from sleep and raised its tip to a high working temperature.The inactivity mechanism is intended to cool the iron when it detects no movement.Source: §6.4 · p. 12

7 Automatic Attack

The paper introduces an automatic method for switching acoustic attack frequencies using real-time inertial sensor data. It demonstrates manipulated navigation orientation, while sample-rate drift limits unattended operation.

7 Automatic Attack

  • The attack program switches between two acoustic frequencies when the induced sensor output drops below a threshold in the opposite direction.The attacker can configure initial frequencies, threshold, and target direction.Source: §7 · p. 12
  • Sample-rate drift limits automatic operation to one or two minutes before an attacker must manually adjust the frequencies.Drift can move the sample-rate multiple outside the interval required for phase pacing.Source: §7 · p. 12
  • An Android application used real-time sensor data to automatically switch attack frequencies and manipulate a smartphone navigation system’s orientation.The demonstration accumulated heading changes of 6.85 rads in 32 seconds and −6.82 rads in about 31 seconds.Source: §7 · pp. 12–13
  • Manipulated navigation orientation can lead users or systems guided by the displayed information onto a wrong path.The passage notes that orientation errors may be harder to calibrate in areas with poor GPS coverage or jammed or spoofed GPS signals.Source: §7 · p. 13

8 Discussion

The discussion outlines design and operational constraints on acoustic attacks, countermeasures, sound-source improvements, and tests showing that lower-frequency signals can also manipulate digitized sensor output.

8.1 Countermeasures

  • Acoustic damping and correctly configured analog filtering address the two identified weaknesses; digital filtering cannot undo aliasing after sampling.Dynamic sample rates could make injected digital-frequency patterns less predictable, while redundant sensors could preserve trustworthy information if one sensor is attacked.Source: §8.1 · p. 14

8.2 Sound Source

  • Eight identical coherent sound sources could theoretically raise sound level by approximately 18.0 dB, increasing possible attack distance by a factor of eight.Practical performance depends on source arrangement, enclosure and horn design, and phase differences.Source: §8.2 · p. 14

8.3 Limitations

  • Moving targets can be difficult to follow and aim at while manually tuning acoustic signals, especially as movement speed and range vary.Possible approaches include predicting trajectories or placing or carrying the sound source near the target.Source: §8.3 · p. 14
  • Manual observation and signal tuning can be slow or ineffective, making timely and correct modulation difficult in some circumstances.More customized devices or automated camera or radar observation are proposed as ways to reduce delays.Source: §8.3 · pp. 14–15
  • Closed-loop behavior can be more complex than the simple signal-mapping model, and user involvement may counter or disrupt attack effects.Attackers may need a more specific model to analyze and predict movement patterns.Source: §8.3 · p. 15

8.4 Generalization

  • Low-frequency signals near the accelerometer’s sample rate manipulated its digitized output without relying on resonance.The Pixel’s ADC sample rate was inferred to be approximately 19.9 Hz; the study also reports that configurable low sample rates can leave nearby signals unfiltered by the analog LPF.Source: §8.4 · p. 15

9 Related Work

The related work spans inertial-sensor attacks, sensor-data eavesdropping, and spoofing of other analog sensors. The paper distinguishes its approach by manipulating oscillating digitized signals to deliver control through non-invasive attacks on embedded inertial sensors.

Attacks on Inertial Sensors

  • Prior inertial-sensor attacks include resonant-sound disruption and accelerometer output attacks, but earlier methods had limits in realism, amplifier requirements, or control over output direction.The paper contrasts its non-invasive embedded-sensor attacks with exposed-sensor biasing and clipping-based control.Source: §9 · pp. 15–16
  • This work manipulates oscillating digitized signals, often treated as noise, to deliver adversarial control through non-invasive attacks on embedded sensors.Prior inertial-sensor work includes a drone crash caused by resonant sound, a sonic-gun demonstration, and accelerometer attacks.Source: §9 · pp. 15–16

Eavesdropping through Inertial Sensors; Analog Sensor Spoofing Attacks

  • Earlier inertial-sensor eavesdropping work recovered keystrokes, speech, or covert channels; this paper instead uses sensor output manipulation for control.Other analog sensor spoofing examples target cardiac devices, infusion pumps, vehicle systems, and UAV optical flow.Source: §9 · p. 16

10 Conclusion

The conclusion presents non-invasive attacks on embedded inertial sensors as a route to implicit control of systems that depend on sensor feedback. It also reports manipulation of oscillating digitized signals using relatively low-frequency signals through the sensing channel.

10 Conclusion

  • Evaluations on 25 devices showed that non-invasive attacks could deliver implicit control to different kinds of systems.Embedded inertial sensors provide feedback used by control systems for critical decisions.Source: §10 · p. 16
  • The work characterized methods for manipulating oscillating digitized signals, which are often regarded as noise, to deliver adversarial control.Source: §10 · p. 16
  • Relatively low-frequency analog signals through the sensing channel could also manipulate the oscillating digitized signal.Source: §10 · p. 16
Loading 1806.07558v2…