Case Studies

The following videos present several case studies performed using MAST with the Carla Simulator. Here, we simulate a mobile ego vehicle and four additional infrastructure sensors. All agents send their detection information to a command center (CC) which shares a global list of detections to each agent. For each case study, a varying number of agents are compromised while the ego vehicle's sensing information is never compromised. The bound boxes are color-coded as follows:

White (ground truth): The ground-truth state of an object with significant overlap with a detection/track. Always pairs with blue.
Blue (true positive): The detection/track state of an object with significant overlap with a ground-truth state. Always pairs with white.
Yellow (false negative): The ground-truth state of an object lacking overlap with a detection/track.
Red (false positive): the detection/track state of an object lacking overlap with a ground-truth state.

Baseline.mp4

Baseline (No Attack)

Compromised agents: none

Uncompromised agents: ego and agents 1,2,3 and 4

As seen in the case study, when the multi-agent sensing system is not under attack, the command center's detections help to improve the sensing performance of the ego vehicle and the other infrastructure sensors.

uncoordinated_1_attacker.mp4

1 Attacker

Compromised agents: agent 1

Uncompromised agents: ego and agents 2,3 and 4

As seen in this case study, agent 1 has added points to its detections. These added points propagate through the command center and affect the ego's shared perception, leading to a potentially hazardous situation.

uncoordinated_2_attacker.mp4

2 Un-coordinated Attackers

Compromised agents: agents 1 and 2

Uncompromised agents: ego, agent 3, and agent 4

In this case study agents 1 and 2 are both inserting fake objects. Here, observe how even more fake objects have propagated through the command center and into the shared perception received by the ego vehicle. Additionally, the fake objects from agent 1 have also been added to the shared perception received by agent 2, leading to even more confusion.

coordinated_2_attack.mp4

2 Coordinated Attackers

Compromised agents: agents 1 and 2

Uncompromised agents: ego, agent 3, and agent 4

In this case study, agent 1 and agent 2 coordinate such that the inserted objects appear in the same place. While the previous uncoordinated attack added more points at random, this attack would be more difficult to detect as both agent 1 and agent 2 are adding an object at the same location. As seen in the video, both the ego vehicle and agent 3 detect the added object via the shared perception from the command module.

uncoordinated_3_attackers.mp4

3 Uncoordinated Attackers

Compromised agents: agents 1, 2, and 3

Uncompromised agents: ego and agent 4

In this final case study, agents 1, 2, and 3 are all adding objects at random. Not only does this severely compromise the ego vehicle's shared perception from the command center, but it further compromises the shared perception of each individual agent as well. This case study demonstrates how even a small number of agents could essentially make the shared perception from the command center useless and lead to potentially dangerous situations.

Page updated

Google Sites

Report abuse