Sequential Attacks on Kalman Filter-based Forward Collision Warning Systems

We use CARLA, a high-fidelity vehicle simulation environment, to generate measurement data that we input to the Kalman filter-based FCW. We took Carla configuration from a publicly-available MATLAB implementation of FCW. The simulation runs at 20 frames/sec and thus, each sensor receives data at that rate. Furthermore, this configuration is commonly available on production vehicles today, and thus, our simulation setup matches real-world FCW systems from a hardware perspective.

We obtain vision and RADAR measurements from the CARLA simulation. Each step in our KF corresponds to one frame of the simulated video sequence (i.e., 0.05 seconds). We perform preprocessing of measurements to remove outliers and interpolate missing data. We assume the KF initializes the distance and velocity prediction by the average of the first vision and RADAR measurement. The acceleration is initialized as 0 for both directions. The covariance matrix is initialized the same as in Matlab FCW.

During braking, the distance between the ego vehicle and the MIO reduces by 12.76m. Since it does not cover the ground-truth distance 14.58m before braking, the collision can be avoided. This validates the effectiveness of FCW.

Our attacker aims at causing a vehicle collision. To accomplish that, the attacker suppresses the first 10 steps of red warning. The ground-truth distance to MIO is 9.58m, which is below the minimum distance needed to avoid collision (12.76m), thus the collision happens.

After 1.2 seconds, the ego vehicle driver realizes they should brake. With sufficient warning from FCW, they can avoid a collision.

We hide the warnings at the beginning. The driver is alerted too late. Thus, a collision cannot be avoided.

We assume there is a trailing vehicle 7m behind the ego vehicle, driving in the same velocity. Our attacker aims at causing the FCW to output red lights, so that the ego vehicle suddenly brakes and results in a rear collision with the trailing vehicle.

To achieve that, the attacker changes 40 green lights to red. Therefore, the driver maintains braking for at least 2 seconds. Assume the driver of the trailing vehicle is distracted, then during the 2 seconds, the distance between the trailing and the ego vehicle reduces by 7.84m>7m, thus the rear-collision happens.

Driving stably in the middle of two vehicles.

The Ego vehicle suddenly brakes. When the driver of the trailing vehicle realizes a collision is about to happen, it is already too late.