IEEE Consumer Electronics Magazine - November/December 2019 - 105

data frames have two essential components for
communicating messages-an arbitration ID
(also known as process ID or PID) field and the
data field. Yet, the mapping for how PIDs and
data fields encode functionality is proprietary
and varies greatly across models.
Thus, the key challenge for state-based methods is mapping the proprietary data fields of the
CAN bus packets to vehicle functions, such as
accelerator, wheel speed, and braking without having documentation of the data formatting. This
motivates development of algorithms for automatic
discovery of classes of functions (e.g., data fields
linearly related to speed) that could be in turn
mapped to states such as acceleration or idling.
State modeling approaches often involve the
use of hidden Markov modeling (HMM) to infer the
state of a system from observables called emissions. For this research, we have chosen to establish a methodology for discovering an HMM that
will support automated decision making processes
in the future. In future endeavors, we believe the
best model for using HMMs to support autonomous decision making and cyberprotection are
partially observable Markov decision process
(POMDP) and predictive state representations
(PSR) as they have been shown to be especially
pertinent for applications involving user (agent)
inputs, environmental awareness and system reactions such as driver-assisted vehicles.
A relational diagram of Markov models (MMs)
is shown in Figure 1, which is an adaptation from
the paper by McAllister and Bui.5 The items at the
bottom left (MM and Markov decision process
[MDP]) represent directly observable states and
decision processes. The upper right pair (PSR
and observable operator models) are models that
incorporate predictive outcomes. A machine
learning approach for identifying the Hierarchical
HMM of an operational vehicle is available.6 The
first step in developing a POMDP for predicting
optimum decisions is to develop multiple levels
of HMMs; hence, we view this paper as a progression toward autonomous decision making.7 The
problem of automatically understanding CAN
messaging is, thus far, only discussed in one reference.8 Their approach requires CAN messages
with ground-truth vehicle function labels (requiring tedious, manual reverse engineering) as training supervised learning techniques.

November/December 2019

Figure 1. Zoubinlike cube showing relationships
among the relevant state-based MMs.

We propose a novel data-analytic approach
for characterizing CAN communications based on
their relationship to physical states of the vehicle. First, we make the critical observation that
the four sequential byte pairs that comprise each
8-b data frame is the appropriate unit of data for
analysis and call these byte pairs data words.
Next, we propose a way to learn speed-related
data words via correlation statistics, which
allows understanding of vehicle speed. This unsupervised discovery of the wheel speeds from the
CAN messages is enough to understand basic
states of the vehicle (e.g., idle, accelerating, etc.)
but is not robust to noise, e.g., a cyberattack
manipulating CAN messages; furthermore, we
seek to understand the other CAN messages in
terms of the vehicle states. Consequently, we
leverage this to represent each data word's values as a linear combination of speed and acceleration. Finally, the groupings inform creation of an
image over small time-windows of the CAN data,
which permits CNN-based image processing techniques to accurately predict the state of the vehicle even with somewhat noisy (maliciously
manipulated) CAN signals. This takes advantage
of work such as described by the paper in Lemley
et al.,9 which supports smart city integration.10

MULTISTAGE STATE AND SYSTEM
DISCOVERY
We have developed a combination of supervised and unsupervised machine learning tools
to extract the relationships between network
data packets and the functions of the vehicle.
The steps for deriving the states of the driver's
actions from CAN bus traffic are described in
Figure 2. The flow begins with the collection of

105
IEEE Consumer Electronics Magazine - November/December 2019

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - November/December 2019
