IEEE Computational Intelligence Magazine - November 2019 - 74

features multiple stand-alone agents that
execute intelligence modeling collaboration without coordination (e.g. swarm
intelligence), but it might also turn into
competition. It is important to note that
additional combinations of such system's models are possible, especially
when they need to meet specific case
studies' specifications.
In today's modern industrial systems,
the intelligence information exchange is
left to application's implementation,
which applies to all the previously mentioned models. For example, when the
learnings of one node is propagated to all
the others in a cooperative model, or
when a central entity issues intelligence
functions to the edges of the system, it is
performed by applications own mechanisms. Also, in the stand-alone case, the
intelligent function's update and configuration at on-boarding time, or during its
operative lifetime, are also handled by
application's implementations. Currently,
in the industrial world, there is no clear
dominant framework or standard that
enables intelligence distribution; however,
there have been attempts in this direction.
For example, the Agent Communication
Language (ACL) from the Foundation
for Intelligent Physical Agents (FIPA)
model was defined as a protocol for communication in multi-agent systems [21].
C. Intelligence Provisioning
Handshake

Intelligence's own nature and sensitivity
requires more protection and handling
care than, for example, displaying a web
page. For intelligence, confidentiality
and accessibility are the main concerns,
therefore a more generic concept of privacy must be accounted for as well. Data
sources are very valuable by themselves,
not only because of the cost of acquiring
high quality training data, but also for
their significance in a larger context and
in conjunction with correlations. For
example, the information related to one
car accident may not be of great concern
for a car manufacturer, but a collection
of data referring to all accidents involving one of their models might be.
Machine learning algorithms' task is
often finding correlations or models that

can be used to make intelligent decisions.
The resulting inferences may be of the
same sensitive nature or even more than
the data sources used by the intelligent
functions. Today, depending on the existing regulatory framework, information
provided or generated by applications
could be exchanged between service
providers and applications with a large
diversity range of detail.
This information flow can be secured
through terms-of-use agreements that
give application users little or no information how the collected data is being
used. Also, these agreements often do not
provide ways for the users to restrict or
discriminate when, what, and with
whom information can be shared. The
user's alternative could be just to quit or
abandon the services, but in the case of
very market dominant applications, it
becomes difficult for users to find suitable fits for their needs. For data sets
of high sensitivities such as customer
records, intelligence service users require
full data privacy coverage even if the data
is processed outside of their premises or
devices. This requirement implies that
intelligence providers should be able to
install machine learning algorithms
without needing to know the actual values utilized as input. This process can be
achieved with measures such as homomorphic encryption [22] or secure multiparty computations [23].
The intelligence providers may be
also concerned about the confidentiality
of the algorithms used from their intelligence harvesting, both from the perspectives of learning processes setup and
recall phase parameters. Providers often
wish to protect their internal knowledge
as well as their research and development investments from intelligence copy
attempts or unauthorized cloning. This
issue becomes even more difficult to
address when the attacks use MI to
clone the knowledge [24], [25] and to
put in place relevant protections is
another area requiring more research.
1) Intelligence Request and
Capabilities Exchange
Configuring and managing a D-MI
requires high flexibility to address the

IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | NOVEMBER 2019

existence of multiple types of devices,
scenarios, services, and topologies of
connectivity. Such flexibility is necessary
from the very starting point when intelligence is first being requested. An initial
exchange of capabilities and relevant
characteristics between the target device
and the intelligence providers allows
providers to assess what type of algorithms and intelligence setup are proper
and relevant for the device according to
the application layer's expected services.
During this exchange, the provider can
evaluate if such provisioning for the
requested intelligence on the target
device is feasible or supported. The re quest process involves negotiation
of data sharing policies and may also
include charging aspects of the intelligence and data provisioning in any
direction. The request can be done
directly to an Intelligence Provider or
through a mediator entity such as an
Intelligence Market implementation.
The market may use part of the in formation contained in the request to
match the intelligence needs requested
to the capabilities and services offered
by the Intelligence Providers available in
the market. This process of requesting
the intelligence, exchanging capabilities,
sharing policies and charging policies
comprises an Intelligence Handshake.
Figure 8 displays an example of
how different ecosystem entities interact
amongst each other in order to enable
intelligence provisioning. The different
procedures are grouped according to
some of the distinct phases of the MI
LCM as defined in Figure 2.
2) Sharing Policies
Sharing policies include what data can
be shared and how, which services can
access which data, and what data requires
additional anonymization. Instead of
being solely left to the parties involved,
data sharing polices should be enforceable by protocols and processes introduced to set up the intelligence services.
Otherwise, there is no warranty that a
change of terms of service agreements
would retroactively affect data exchanged
under different agreed conditions. Profiles can be provisioned in multiple ways

IEEE Computational Intelligence Magazine - November 2019

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - November 2019