Page 87 - RAC_CIAW_ a_I_n_01_2021.pdf

nates of an individual j of the optimization algorithm, way, attack solutions with greater amounts of packet
wherein [W ct, j , W ] are estimations of [W , W ]. Let loss are more distant from being chosen by the optimi-
ct
fb
fb, j
ŷ ( ) denote the plant output of an NCS simulated by zation algorithm. Therefore, to find the most suitable
j
the optimization algorithm, in which the set of frames attack solution, the optimization algorithm must mini-
outlined by x = [W ct, j , W ] are lost. It is assumed that mize the objective function (5).
j
fb, j
the control function C (z), the plant transfer function
P (z), and the communication delays of the real NCS, RESULTS
which are used in the simulations to obtain ŷ ( ), are
j
known by the attacker (obtained, for instance, through This section evaluates the optimized data loss at-
a System Identification attack [14]). tack proposed in Section III and compares its perfor-
mance with the original attack described in [7]. The
Thereby, to produce an optimized attack, capa-
ble of causing accurate overshoots in the plant with results were obtained through simulations using MA-
a reduced number of dropped frames, it is proposed TLAB/SIMULINK. The source code used throughout
here the objective function ℱ computed according to the simulations was exactly the same as the original
j implementation. For the sake of comparison, the atta-
(5). Such function, used to compute the fitness of each
individual j of the optimization algorithm, includes cked NCS is the same as the one used in [7], consisting
a new term (described below) designed to reduce the of a DC motor, controlled by a proportional-integral
number of frames that the attacker will cause the loss. (PI) algorithm, with a sampling time of 20 ms. For
more details on the PI control function C(z), the plant
transfer function P(z) and all other information about
the NCS configuration, we refer the reader to [7].
As described in Section III, the new attack stra-
tegy proposed in this paper is implemented using two
alternatives of optimization algorithms: the PSO and
the BSA. The parameters of the PSO are ω = 1, φ =
1
φ = 2 and δ = 1. The BSA is configured with = 1.
2
Both algorithms are configured with a population of
100 individuals and are executed for 600 iterations.
The parameters of the objective function (5) are the
same as those used [7]: Υ = 1.5 / , = 10, = 30,
1 2
= 10000, ρ = 0.5 / , = 100, = 200, and
The first term ℱ aims to make the plant output = 1 / . Each time that the attack is executed, both
l
ss

1, j
during the overshoot reach the peak level intended by sequences and have h=50 frames each, beginning
the attacker, defined as Υ. Additionally, it specifies that when the DC motor starts up.
fb
ct
the instant when the peak of the overshoot occurs peak ,
needs to be within a time interval bounded by and k 1 For the sake of comparison, three scenarios are
and . Otherwise, the individual is penalized by , an considered: a scenario with the original BSA-based at-
2
empirically defined value. The second term ℱ aims to tack design [7]; and two scenarios with the improved
2, j
prevent that after peak (after the overshoot) the output attack herein proposed. One of these two is implemen-
of the plant ŷ ( ) does not assume values below the ted based on the PSO and the other based on the BSA.
j
limit r stipulated by the attacker. If a level below r is For each simulation scenario, the words [W , W ]
ct
fb
reached, the individual is penalized by . The third consist of 200 network packets each. To statistically
term ℱ aims to ensure that after a moment , prede- assess the performance of the three attack implementa-
s
3, j
fined by the attacker, the plant output converges to the tions, 100 simulations are executed for each scenario.
stationary value normally assumed when it is not Fig. 2 shows the mean of the plant’s output signals
ss
under attack. Finally, the fourth term ℱ , introduced y ( ) for each attack implementation. Each mean signal
4, j
in this paper, adds to the fitness function the inverse of is computed based on the results obtained over the afo-
total number of dropped frames (L W ct, j + L W fb, j ), whe- rementioned 100 simulations for each scenario. Also,
rein L W ct, j and L W fb, j are the total number of frames lost Table I presents the mean and the standard deviation
in the control and feedback signals, respectively. In this of the overshoot’s peak achieved in each attack im-
CIAW – EFICIÊNCIA, CULTURA E TRADIÇÃO 87