IEEE Consumer Electronics Magazine - May/June 2022 - 34

AI Security for MEC
based on a rule-based detection technique. The
rule-based detection technique, executed at the
generator system, corresponds to a set of attack
signatures defined by the security expert, which
are updated over time and depend on the number
of suspected behaviors that are identified by the
experts as possible attacks. An example of an
attack signature could be the number of packets
sent and dropped that are above certain thresholds
to detect the DoS attack, where the packets
sent and dropped are the relevant features. A set
of attack signatures are stored in the signatures
database of the generator.
In this cyber security context, the generator
and discriminators cooperate between each other
to increase the attack detection rate and reduce
considerably the number of false positives. This
cooperative detection is illustrated in Figure 2. The
generator sends to the discriminators a list of
attacks to be detected (with the related signatures)
and then each discriminator extracts its
required information. Specifically, in the preprocessing
phase, ATDD extracts the values of features
from the signatures and associates them
with their corresponding labels while ANDD
extracts only the values of features from the signatures
without associating them with their labels.
Afterward, discriminators execute the training process
by using these feature values as inputs. ATDD
runs a supervised multiclass deep learning algorithm13
that builds the normal and attack patterns
during the training process and then classifies the
new incoming data as an attack or normal behavior
during the classification/detection process according
to the patterns determined in the training process.
The normal and attack patterns obtained
during the training process correspond to clusters
of vectors related to each pattern. ANDD uses as
inputs the outputs of ATDDand runs a deepconvolutional
GAN.14 This discriminator also analyzes
the network traffic to detect anomalies. If an anomaly
is detected, ANDD first checks if it is also
detected by ATDD. In the case of this anomaly only
detected by ANDD, an alert message is sent to the
security expert in order to update the attack signature
data base, i.e., adding new attack signatures.
The alert message includes attack features along
with the type of attack. Note that the cyber security
expert feeds the generator periodically with
new attack signatures in order to improve the
34
training process and hence to increase the attack
detection rate. In addition, the security expert
investigates the detection accuracy of ATDD
against zero-day attacks with the goal to decrease
the false positive rate.
Stackelberg Trust Game
The hybrid learning security engines based on
the GAN approach embedded in MEC servers
cooperate with each other by exchanging their
detection decisions (i.e., attacks with the related
signatures and features) in order to detect accurately
the new category of attacks, i.e., zero-day
attacks. However, a hybrid learning security
engine could be infected by the attacks and the
infected security engine could convey false detection
decisions to its neighboring engines; therefore,
the accuracy detection of zero-day attacks is
impacted, i.e., the increase and decrease of the
false positive and detection rates, respectively.
Thereby, to overcome this security issue, each
hybrid learning security engine should monitor
the trustworthiness of the detection decisions provided
by its neighboring security engines with
whom it collaborates. The interaction between
security engines is modeled as a Stackelberg security
game, where the hybrid learning security
engine is the leader player and its neighboring
security engines are the follower players. In the
security game, we assume that almost all follower
players are selfish and aim to impact the attack
detection of the leader player by providing false
detection decisions. In this noncooperative game,
the leader player aims to maximize its utility with
consideration of the best strategies undertaken by
the follower players and vice-versa. In the Stackelberg
game, the expected utility functions of the
noncooperative players depend mainly on the
number of malicious hybrid-learning security
engines that the leader player detects and the
number of false detections that the leader player
generates against the malicious follower players.
In the proposed Stackelberg security game, the
leader player launches its optimal strategy for
detecting the malicious hybrid learning security
engine, by considering the best response of the
follower's strategy. Furthermore, the malicious
security engine executes its optimal strategy for
providing a false detection decision, by considering
the best response of the leader's strategy. Note
IEEE Consumer Electronics Magazine
IEEE Consumer Electronics Magazine - May/June 2022

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - May/June 2022
