The modern-day vehicle’s driverless or driver-assisted systems are developed by
sensing the surroundings using a combination of camera, lidar, and other related
sensors by forming an accurate perception of the driving environment. Machine
learning algorithms help in forming perception and perform planning and control
of the vehicle. The control of the vehicle which reflects safety depends on the
accurate understanding of the surroundings by the trained machine learning
models by subdividing a camera image fed into multiple segments or objects. The
semantic segmentation system comes with the objective of assigning predefined
class labels such as tree, road, and the like to each pixel of an image. Any
security attacks on pixel classification nodes of the segmentation systems based
on deep learning result in the failure of the driver assistance or autonomous
vehicle safety functionalities due to a falsely formed perception. The security
compromisations on the pixel classification head of the object segmentation
systems result in falsely segmented pixels from the incoming camera images by
corrupted pixel labels with wrong object classes for the pixels. The popular
encoder–decoder-based deep learning object segmentation network is considered,
which is vulnerable to these attacks in its last fully connected neural network
layer. Hence, the cryptographic solution mechanism is proposed here, where the
pixel classes are encrypted and signed in the classification network nodes
before applying the activation functions. RSA-512 algorithm-based encryption and
DSA-512 algorithm-based digital signature are used to generate the proposed
cryptographic components. The added cryptographic components are verified upon
segmenting the objects to ensure the segmented object information is free from
described security attacks. The performance of the proposed cryptographic secure
object segmentation is evaluated for the popular segmentation network called
U-Net for the Cityscapes segmentation dataset with the proposed cryptographic
algorithms. The performance evaluation indicates that the secure semantic
segmentation is performed with satisfactory precision, recall, and F1 scores of
0.86, 0.85, and 0.85, respectively, along with the added security
components.
