Moving Case Studies

Experimental Setup

For the following case studies, the attacking vehicle starts out in front of the victim's vehicle. The mmRAD system is placed in the trunk of the attacking vehicle so that it can sense the victim's radar parameters and launch attacks. At the start of the attack, the attacker and victim begin driving forward at 30 mph and 10 mph, respectively.

When reviewing the following results, we make the following notes:

The minimum detection range of the victim radar is 24.35 m*.
Detections corresponding to stationary vehicles and objects are not shown in the plots below**.

About Moving Case Studies

When driving, automotive radar sensors are used my many safety-critical systems including forward collision warning (FCW) and blind spot monitoring (BSM) system. Moving forward, radars are being utilized for various advanced driver assistance systems (ADAS) including automatic cruise control (ACC) and other autonomous driving applications. Regardless of the application, it is critical that the radars in these systems accurately sense their environment. In the following case study, we demonstrate the first known black-box translation attack where an attacker is effectively able to "move" objects in the victim radar's point cloud. To provide a point of comparison, we also present a similar case study without such an attack as well. Finally, the figure on the right features the experimental setup used for these case studies.

No Attack

Translation Attack

No_Attack_Combined_revA.mp4

Case Study Movie

No Attack

Here we demonstrate the baseline performance of the victim radar detecting objects moving along the road. This demonstrates the radar's performance when no vehicle is present. For this case study, the "victim" vehicle was moving at roughly 10 mph, while the black was moving at a speed of roughly 25 mph.

Detected Object Locations

This figure documents the range that the victim radar perceived the target to be for each frame of the case study. When not under attack, the victim radar detects the target moving away from it in a majority of its frames

Attack Timeline

The figure above documents the progression of the case study. Here, the top row shows the Range-Doppler response while the bottom row shows the CA-CFAR detections. See how the target's reflection can clearly be seen in the victim's Range-Doppler response and CFAR detections when no attack is present. Additionally, we have highlighted how the radar picks up "ground clutter" which is the result of reflections from all of the objects that aren't moving in the scene (i.e: parked cars, sidewalks, trees, etc.)

Translation_Combined.mp4

Attack Movie

Translation Attack

In this case study, we perform the first case study of a black-box translation attack on a moving platform. To accomplish this, we start with a false-negative attack centered at a range of 75m and propagating away from the radar with a velocity of 10m/s. Then we simultaneously launch a false positive attack starting at 100m and propagating towards the radar with a velocity of 12m/s. The resulting attack leads the victim radar to mistakenly believe that the attacking vehicle is heading towards it when it is actually heading away from it. Observe how the victim is only able to detect the attacker vehicle once the attack finishes. This case study is a crucial demonstration as a malicious actor could utilize this same method to trigger a dangerous and unexpected braking maneuver leading to a severe collision on the highway.

Attack Overview

Detected Object Locations

This figure documents the range that the victim radar perceived the target to be for each frame of the case study. Here, the green dashed lines represent the actual target locations while the blue dots represent the location detected by the victim radar. The case study begins with the attacker launching a translation attack where we highlight that the victim mistakenly detects the attacker's vehicle as coming towards it, potentially leading to an emergency braking maneuver. Once the attack finishes, the victim detects the actual location of the attacker.

Attack Timeline

The figure above documents the progression of the case study. Here, the top row shows the Range-Doppler response while the bottom row shows the CA-CFAR detections. The first and second columns show the victim's radar perception when not under attack while the third column represents the radar's perception when under attack. When no attack is present, we clearly see that the attacker vehicle is detected in the Range-Doppler Response and the CFAR detector. However, when the attack is engaged , we see that the victim fails to detect the attacker vehicle and instead detects the spoofed point added by the attacker.

Footnotes:

*This is a result of the CA-CFAR detector configuration (see Sec.VII in the paper for a more detailed explanation) and hardware limitations. Thus, it is possible that some targets at close range will not be detected until they move further away.

**Stationary vehicles are significantly more difficult to detect as they often get lost in the clutter from the surrounding environment. Additionally, these detections are not shown in order to make it easier to observe the behavior of targets of interest.

Page updated

Google Sites

Report abuse