IEEE Electrification - March 2021 - 43

rejecting DR calls or committing the DR capacity. Tampering with this information can have severe effects
on the effectiveness of the DR program and power
grid operation. For example, false data injection
attacks (FDIAs) on the information sent to customers
inherently forces them to make suboptimal decisions
on how to use their appliances. Similarly, FDIAs on the
information sent to the utility/aggregator (e.g., accepting/rejecting the DR calls, SMs data) misinforms them
and, hence, forces their DR scheduling routines and
algorithm to produce erroneous dispatch decisions.
These attacks are considered as causative attacks on
decision-support and learning schemes employed by
the utility/aggregator. Such false DR schedules and
customer-end responses can incur operational challenges, such as frequency and voltage excursions, and
increase the system operating cost due to a mismatch
between the committed DR capacity and the DR
capacity provided in real time.
3)	Availability: This implies that any authorized entity is
deprived of reliable and timely access to services and
information. The effectiveness of DR programs relies
on the uninterrupted and timely exchange of information between the utility/aggregator and customers.
In turn, disrupting this information exchange by
exploiting customer-end devices (such as SMs and
smartphones), the communication channel between
the utility/aggregator and customers, and the DRAS
server can damage the efficacy of the DR program. For
example, denial-of-service (DoS) attacks on SM data or
on the responses of customers to DR calls can inject
erroneous values into the training data used by the
learning algorithm deployed by the utility/aggregator.
This attack misleads the algorithm to design suboptimal DR schedules. Similarly, DoS attacks on DR schedules, DRAS servers, or VENs preclude DR customers
from participating in DR calls. This may, in turn,
undermine the trustworthiness of the DR program
deployed by the utility/aggregator.
The ability of adversaries to compromise customer-end
devices, such as SMs and smartphones, utility/aggregator
DRAS, and DR communication channels, has been greatly
aided by the automation of DR programs and by a lack of
standardization of these programs across the industry. For
example, there is no internationally (or, in the case of the
United States, interstate) accepted DR communication
protocol. Although some protocols (e.g., OpenADR 2.0)
have recently gained acceptance, they are still not recognized at the regulatory level. The OpenADR 2.0 protocol
authenticates, encrypts, and digitally signs the DR information exchanged between the two parties. Although
utilities use the OpenADR protocol, the aggregator may
use a proprietary communication protocol whose security remains undetermined.
Recently, blockchain techniques have been validated
as promising solutions to avoiding centralized

authentication and data storage. The use of blockchain
schemes inherently increases the security and privacy of
DR customers by means of: 1) a trustless decentralized
network, 2) immutability, and 3) network consensus.
Unlike the centralized security authority, each blockchain is managed by an anonymized decentralized node
(SMs or VENs in DR) that verifies the authenticity of new
nodes and data using network consensus. Furthermore,
blockchains maintain a timestamped and hashed list of
data. For example, SM data can be saved and communicated in blocks with a hash that depends on previous
blocks. This hashing increases the difficulty for an
attacker to tamper with data. This feature will allow SMs
to authenticate themselves and encrypt data without the
use of DRAS. Although blockchains have more-robust
security features compared to centralized security mechanisms, they suffer from cryptojacking-where adversaries can access unauthorized computations across
blockchain nodes-thwarting genuine transactions and
leading to system failures.
Even with standard security mechanisms, such as
OpenADR and decentralized blockchain deployed by the
utility and aggregator, DR security ultimately hinges on
the cyberhygiene of DR customers. This is particularly
concerning as many customers are either unaware of the
needed security measures (e.g., strong passwords and
multifactor authentication for their IoT devices and applications) or cannot afford these security measures (e.g.,
SMs and BEMS with a high computation power for industry-grade encryption).
Given these cyber vulnerabilities and proven attack
mechanisms, ML algorithms, such as online or the Q/Zlearning discussed in the " Online Learning Processes " and
" MDPs " sections, respectively, can improve the security of
DR programs. For example, the online learning algorithm
executed on the DRAS updates the incentive signals sent
to DR participants during a real-time DR event. Unlike an
offline approach, where all of the historical data are used
to analyze the sensitivity of the DR customer to a price
signal during a given DR event, the online counterpart
only uses current DR event data. This bounds the effect of
integrity attacks on training data. Similarly, Q/Z-learning
can return probabilistic estimates on the DR participant's
power curtailment, given historical DR incentives and
operating conditions, such as the temperature for ACs,
without necessarily monitoring SM data in every DR
event. Unlike conventional DR approaches, where SM data
are the backbone, ML algorithms, such as conditional kernel density forecast, learn the customer-end energy usage
without requiring (or at least reducing) the SM measurement, which reduces data dependency and, thus, an
attack surface. Furthermore, customer-end responses sent
to the utility/aggregator can also be estimated based on
their historical responses. Although these techniques
decrease the attack surface arising from a lack of customer-end cyberhygiene, they are not a panacea.
	

IEEE Electrific ation Magazine / MARCH 2 0 2 1

43
IEEE Electrification - March 2021

Table of Contents for the Digital Edition of IEEE Electrification - March 2021
