Risk Analysis of Terrorist Attacks on Seaport Facilities Using Bayesian Networks

Risk definition and analysis in ports

Different forms of risk concepts are proposed in the risk sector, and each measures risk from different perspectives (Yu et al., 2020). Multiplying P and C is one of the most formal and well-established definitions of risk (R), where P is the probability of risk and C is the consequence (Rausand, 2013). However, in addition to the P and C, the risk is a complicated notion, and other variety of variables, such as uncertainty, exposure, and scenarios, are also included (Aven, 2012). For instance, Kaplan & Garrick (1981) add Specific Scenario (s), and Aven (2012) suggests considering Background Knowledge (BK), and Uncertainty of Data (U).

Within this context, it is suggested that R is to be measured through proof and observations following the Bayes risk theory. As opposed to the normal probabilistic theory, the Bayes risk theory models BK and U by using factors of earlier probability and conditional probability to illustrate a frequency probability and the interdependency among the risk components. As a result, risk should be updated to a model with a set of prior data and conditional probabilities as it is a dynamic notion that changes depending on the circumstances when new data is added to the probability model (Yu et al., 2020).

In fact, the port industry is a high-risk, high-return venture. The situation is made significantly worse by the emerging security risk of maritime terrorism toward ports. Due to its nature of randomness and scarcity in historical failure data, TAs are among the most unpredictable threats that can happen at both the sea and land sides of a port.

Terrorists who committed unlawful violence in a nautical environment, against ships or permanent platforms at sea or in ports, any passengers or crew aboard ships, and coastal infrastructure or centres involving resorts for tourists, port locations, and port cities are referred to as maritime terrorist (Bergqvist, 2014). In order to calculate risk posed by terrorists, NRC Committee has proposed a model from a terrorist perspective (Ezell et al., 2010). This model however needs to satisfy two important assumptions, which are 1) it assumes ideal adversary intelligence and rationality and 2) the intelligence community knows the objectives the adversary is trying to maximize in the first place. To complement the current state of the art in anti-terrorism risk studies, a new risk analysis model that can take into different stakeholder perspectives on the M-T pairs and their associated risk levels is highly demanded and beneficial.

Bayesian networks

BNs are visual models that combine probability theory with graph theory and it can compensate for the lack of secondary data or insufficient information, it is capable to mix several bits of information and employ expert judgement. An approach to solve sophisticated issues involving uncertainties in input data is provided by BN. BN also can combining simpler and smaller models, can formulate and present a complex system. BN can be used to make various prediction, diagnosis, and it can model the interdependencies among the root causes which cannot be achieved by simple hierarchical approaches such as fault tree analysis (Abdul Halim, 2020). Fault tree analysis reveals an either/or possibility which only presents the event that can occur or not (it is either one or zero probabilities). However, BN can model the partial occurrence of event B when event A occurs.

The probability of nodes, which is described as a means of communicating information or conviction that an event will occur or has occurred, is a characteristic of BN (Abdul Halim, 2020). The Bayesian model allows relevant information which acts as influencing factors to be incorporated. The deeper knowledge of the parameter failure rate that has resulted from the inclusion of these influencing factors, has increased trust in the analysis of the overall results. It is recommended that a BN model can be deployed using the prior information gathered for each influencing parameter together with the advice of experts (Jones et al., 2010). The methods make BN a trustworthy model for probabilistic inference by allowing the influence of evidence regarding one node, facilitating the propagation of other nodes in multiple-connected trees. Due to such advantages, BNs have been widely applied in maritime/oceanic risk analysis (Chang et al., 2020; Yang et al., 2018), but in port yet (e.g., John et al., 2016).

Root cause analysis (RCA)

RCA refers to a specific underlying cause that can be identified, fixed and prevent recurrences can be generated (Rooney & Heuvel, 2004). RCA is used for fault localization, fault isolation or alarm/event correlation. It is the method of inferring the set of faults that cause a given set of symptoms. If the root cause/fault are directly observable, this process could be trivial, in which case they are also symptoms. However, in complex systems this is not the usual case. When there were one or more factors may actually constitute the root cause(s) of the problem being researched, it is established practice to refer to the root cause in a singular form. In this study, the procedure of identifying BN nodes is carried out using RCA. This strategy explains how an attack event could occur in container terminals owing to ambiguous conditions while also seeking a technique that can be utilized to describe causality (Rooney & Heuvel, 2004).

In the model, different types of possible TAs are explored and identified with respect to different port sites and facilities through a thorough literature review and secondary data analysis. Additionally, the brainstorm technique is used for facilitating the discussion with experts for subjective risk data elicitation. The RCA method is used in this process to explain the M-T pair identification and the interdependent nodes toward operation field experts. The investigation and mitigation of risks should be done in an organised manner and structured approach, but there are issues with RCA, where many RCAs are performed improperly or insufficiently and do not yield helpful data. Organizations frequently address each RCA separately, rather than drawing conclusions as a whole (Wu et al., 2008). Therefore, in this paper, we propose a holistic BN-RCA approach to take advantages of both methods. Specifically, it allows RCA to aid the identification of all the root causes, and a BN method to model the interrelationship among the risk factors for the risk prioritisation of the identified M-T pairs.

See Full Size >

Step 1: Historical data acquiring, fitting and classifying

This step helps identify TA modes and targeted port sites/facilities (M-T) using historical data from Global Terrorist Database (GTD). This database shows a list of maritime TAs that cover the accidents at both sea and ports since 2001. After extracting port cases (including those in port waters), 55 cases were collected (see Appendix 1) and distributed in Figure 2 by years. It reveals that the overall tendency on terrorist accidents in ports is growing despite variation, which further illustrates the urgency and significance of this research work in practice.

See Full Size >

Step 2: Filtering repetitive cases and listing frequent scenarios

Based on this database, all 55 cases were analysed and filtered. Every repetitive case will be put together in a group and a list of frequent scenarios were generated. In order to confirm that the scenarios accurately reflect reality, expert opinions were sought. A selected expert denoted an individual with 5 to 20 years of involvement in the maritime security field. In this case, six experts were selected where three of them are from the port police department and another three from the rescue department. Specifically, a judgement involved weighing available proof and drawing a holistic conclusion. In the study context, an expert role provided the judgements based on the experiences involved in making sound evaluations.

Step 3: Identify the cause nodes

This step has been briefed extensively in Wan’s thesis which shows how researcher uses the RCA method to identify the cause node, verified it and classified it into primary causes and secondary causes (Abdul Halim, 2020). By doing so, the attacking modes (M) and the vulnerability port facilities (T) in the first two objectives can be met.

Step 4: Define the discrete states of the variables (Nodes)

Binary nodes are used in this initiating work in the field to simply the subjective data elicitation from domain experts (Wan, 2020). Therefore, two axioms were incorporated to be acceptable under the BN algorithm:

•For any event, 0 ≤ P (X=x) ≤ 1, with P (X=x) = 1 if and only if X=x happens with certainty.

•For any two mutually exclusive events X=x and X=y, the probability that either X=x or X=y occurs is P(X=x or X=y) = P(X=x) + P(X=y).

Step 5: Developing a BN Model

Confirming their relationships and building BN model based on their relationships (Abdul Halim, 2020).

Step 6: Check and modify the model by using a D-separation technique

Using D-separation technique to investigate the correctness of the network in this methodology (Wan, 2020).

Step 7: Data collection and analysis of each node

This paper uses both qualitative and quantitative data which can be into 3 steps mainly 1) collect data under the assumption, 2) qualitative data collected from experts 3) validation. Quantitative data is any data that can be counted or expressed numerically while addressing questions pertaining to quantity, frequency, value, or size such as the news report on the TA and the GTD that provided historical data on past maritime-oriented TAs. The qualitative and quantitative data are very different, but with both of them, a complete picture of an event can be captured. Data that resembles something or characterises an item or phenomenon is referred to as qualitative data. It is subjective, investigative, and focused on comprehension of a problem or situation and depends on descriptive words, images, and observations (Davidson, 2019) for example by interviewing several experts.

Collecting data under assumption

It was impossible to predict the behaviour of terrorists since they sought information and exploited weaknesses in defences to increase the impact of attacks (Abdul Halim, 2020). Therefore, to ensure consistent feedback from different experts, it is essential in an M-T scenario analysis, to set a condition that the terrorists have already set their T in an investigated port. As a result, in this paper, we assume an attack occurs by assigning a 100% prior probability across different root causes, and it will be helpful to prioritise the high-risk level M-Ts effectively.

Furthermore, setting the condition is to make sure there will be commonality in the scene for the experts and hence the received results from different experts are presented in a common plate and hence tend to consistent. It is simply because that the different setting leads to different interpretations of the TAs on a container port terminal. Furthermore, the data collection subject to an impression of imminent attacks facilitates the risk prioritisation.

Qualitative data calculation

This step has been briefed extensively in Wan’s thesis which shows how researchers calculate the qualitative data from experts by instructing them to fill up the range of scales in table 1 in responding for each of the situations given during the interview (Abdul Halim, 2020).

Those responses and answers collected from experts then were converted into a probability value and then averaged by following the formula below:

$\left(\begin{array}{c}Average\\ probability\mathrm{}value\\ for\mathrm{}a\mathrm{}state\end{array}\right)=\frac{\left(\begin{array}{c}Total\mathrm{}Probability\mathrm{}rate\mathrm{}given\mathrm{}by\\ expert\mathrm{}for\mathrm{}the\mathrm{}same\mathrm{}state/event\end{array}\right)}{Total\mathrm{}total\mathrm{}number\mathrm{}of\mathrm{}expert}$ (1)

Validation

The model's output must at the very least conform to the following two axioms if the model is valid and its rationale is logical (Abdul Halim, 2020):

Axiom 1. A slight increment in the degrees of belief that an attack will happen should certainly result in the effect of a relative increment in the degrees of belief of the Risk of facilities being attacked.

Axiom 2. The total influence magnitudes of the combination of the probability variations from x attributes (evidence) on the values should be always greater than the one from the set of x – y (y ∈ x) attributes (sub-evidence).

Step 8: Bayesian Inference

This step has been briefed extensively in Wan’s thesis which shows how researcher uses Bayes’ Theorem in two ways 1) prior and 2) posterior and then use it in the NETICA Software tool (Abdul Halim, 2020).

Step 9: Sensitivity Analysis

This step has been briefed extensively in Wan’s thesis which shows how sensitivity analysis is important to prove the robustness of the BN model (Abdul Halim, 2020). Parameter sensitivity typically consists of a number of exams/tests in which to examine how a change in the parameter causes alters the model, the modeler sets several/different parameter values. By examining the uncertainties that are frequently connected to model parameters, sensitivity analysis also helps increase trust in the model (Breierova & Choudari, 2001). By doing so, the third objective of this study can be achieved, while insightful implications can be drawn appropriately.

Case study

Step 1: Historical Data Acquiring, Fitting and Classifying

This step has been briefed extensively in Wan’s thesis which shows how the researcher has acquired the historical data from Global Terrorism Database (Wan, 2020).

Step 2: Filtering Repetitive Cases and Listing Frequent Scenarios

Based on this database and the analysis of repetitive cases, four representative attack scenarios are generated. After the expert consultation about the four scenarios, they are verified as 1) terrorists to directly bomb the port, 2) to ramp the port gate or wharf, 3) to attack using weapons from outside and 4) internal attacks using forged employee identities. This also reflects in part, the findings from a previous maritime security study (i.e., Yang et al., 2009a), where the above numbers 1 and 4 scenarios were identified and analysed (Wan 2020).

Step 3: Identify Significant Influencing Risk Variables/Nodes

This step has been briefed extensively in Wan’s thesis which shows how the researcher sort out all the cases of terrorist attack that only happen in the port area, identify on how the whole sequence of attack happened and then list it as the terrorist attack mode on port (Abdul Halim, 2020): -

• Terrorists attacked the port using a truck filled with explosives.
• Terrorists hijacked a vessel and then use it to attack the port
• Terrorists smuggling an explosive in to the port compound
• Terrorists infiltrate the port as a worker
• Terrorists infiltrate the port as a visitor/outsider contractor

From that list, 19 nodes were derived which shows that the attacks may come from the sea and from the land and these nodes then were discussed and verified by the expert (Abdul Halim, 2020): -

1. Using Tampered Truck(s)

2. Hijack Using Vessel(s)

3. Overcome the Prevention of Unauthorized Entry

4. Suicide Collision by Trucks/Vessels

5. Using Tampered Containers

6. Overcome Identification of Employees

7. Overcome the Prevention of Unauthorized Document Access

8. Smuggling Unauthorized Containers (Bombs)

9. Overcome Routine Security Inspections

10. Container Bomb Attacks

11. Overcome Identification of Visitors

12. Overcome the Prevention of Unauthorized Introduction of Items into Port Facilities

13. Armed Attackers Overcoming the Prevention of Unauthorized Entry

14. Weapon Attacks

15. Port Gates

16. Wharf Operation Sites

17. Yard Operation Sites

18. Administration Sites

19. Security Level

Step 4: Define Discrete States of the Variables (Nodes)

This step has been briefed extensively in Wan’s thesis which shows how the researcher has assigned states on each node as shown in the Table 2 (Abdul Halim, 2020).

Step 5: Developing a BN model

A top-down approach was used in developing the model (See Figure 3) with the assistance of RCA in Step 3, starting with an attack mode (M), through an attack target/place (T) and ending at the goal node (SL) (Abdul Halim, 2020).

See Full Size >

Step 6: Check and verify the model by using a d-separation technique

This step has been briefed extensively in Wan’s thesis which shows how the researcher has used d-separation technique to test the relationship of each node and a few modifications were made (Abdul Halim, 2020). Thus, Suicide Collision by Trucks/Vessels nodes was converted into two nodes, which is Suicide Collision by Trucks and Suicide Collision by Vessels

Step 7: Data collection and probability analysis of each node

Unconditional Probability

Five Unconditional Probabilities Tables (UCPTs) are obtained by using historical data from GTD. Among the 55 identified cases, there are 6 UTT related, 13 HVU, 15 UTC, 1 OIE, and 20 OIV. The unconditional probabilities of the give root nodes belonging to the state of “Yes” are calculated as 10.9%, 23.4%, 27.3%, 1.82% and 36.4%, accordingly, as seen in Table 3.

Results from the interviews and questionnaires

Six experts (from port security and maritime departments) undertook a survey. The data were then inserted in the Bayesian model created by using NETICA and then the result was generated. 15 Conditional Probabilities Tables (CPT) are obtained based on expert judgements and presented in Table 4.

The CPTs of the goal node referring to the security level are presented in Table 5, where the different given states are “Low Security” and “High Security”.

Step 8: Bayesian Inference

Using the Baye’s Theorem and the associated joint and marginal probability interference, the developed BN security model in Figure 4 together with the conditional probabilities analysed in Step 7, can be used to formulate a quantitative BN model for container terminal security level analysis. Here, a commercial software package Netica is used to facilitate the computation.

The results demonstrate that the posterior probability value of the node "WA-NO" increases from 68.8% (Figure 4) to 84.3% (Figure 5) after providing a piece of evidence to the node "OPUEARMY absolute NO" in Figure 7, hence the probability of “SL-High” increases from 82.8% (Figure 4) to 89.4% (Figure 5). It means that when WA (weapon attacks) becomes impossible in an investigated port, the port security level (belong to the state “High”) will increase.

See Full Size >

See Full Size >

Step 9: Model Validation and Sensitivity Analysis

To guarantee its dependability, the produced BN is validated by a two-step validation process. The validation framework contains two conceptual tests of a face validity and content validity (Yu et al., 2020). The face validity tests the BN through a face investigation on the BN structures to improve the confidence of the developed relationships, whereas the content validity assesses the content of the BN can be accepted and consistent with reality (Pitchforth & Mengersen, 2013).

Face validity

By contrasting the correlations in the model with the expert's background knowledge and prior research, the face validity of the BN's rationality is validated. The expert panel introduced in Section 3 is invited to evaluate the model consistency. Not only do the elements utilised in the model, which cover every conceivable M-T pairs that affect TAs in ports, show excellent agreement with reality; the interactions between the factors also exhibit excellent consistency with the knowledge of the expert. It is an additional assurance to the model’s robustness beside the d-separation and RCA. As a result, the BN is recognised in the face validity and qualified to deliver accurate simulation on the TAs in seaports.

Content validity

The content validity tries to discuss if the BN's results are accurate. To lead to further risk mitigations, the most essential factor should be chosen. Plus, the importance of the selected factor should meet the human sense.To prioritise the variables, an entropy-based sensitivity analysis technique (for example mutual information analysis) is used. Here, we establish that a high entropy factor is more illuminating/informative than a low entropy factor. With the help of the Netica programme, the mutual information entropy for each node was determined, and the results are displayed in Table 6. It should be emphasised that the node of SL is chosen as the target to compare the relative importance between the target node and the related node.

The SL target node receives the greatest entropy value and variance among all the factors, with values of 0.65941 and 0.14159 respectively. With an entropy value of 0.18865 and relative relevance of 28.6%, the WA is determined to be the element that has the greatest impact on the SL. It followed by WOS, YOS and AS, which are three highly influential factors to the SL. The content validity shows the developed model follows Axiom 1 that was introduced in Section 3 (i.e. Step 7), thus validating the BN is rational and logical. To analyse the impact of each pair of M-T on the port security level, the results are obtained and presented in Table 7.

From Table 7, it reveals that the most influencing M-T pairs are Suicide Collisions by Trucks/Vessels in a decrement order, which obtains a severity level of 34.6%. The model shows two nodes (i.e. UTT and PG) are changed due to the potential influence from the incident. When the incident happens (i.e. trucks/vessels is 100% yes), the probability for using tempered trucks raise from 10.9% to 100% and the probability for port gates also significantly increase from 5.9% to 76%. It means that effective security risk control measures should be developed with respect to this priority list for the most effective risk reduction. In the meantime, it is noted the incident of smuggling unauthorized containers not only shows high risk (i.e. 26.9%) but also leads to wide fluctuations of seven nodes changes in the model. Comparing the abovementioned two M-T pairs, we find that some incidents like suicide collision can be defined as a typical direct incident, which has a short and simple causation chain but leads to high risk. In contrast, the causation chains for the incidents like smuggling unauthorised containers are more complicated, which has a wider range of influence and the state variations are slight and difficult to be detected. These indirect incidents are more easily ignored, that requires more attention in daily management. This finding can effectively help improve security risk control by focusing on direct incidents. In comparison with the results of single node evidence (i.e. x-y evidence in Axiom 2) the ones provide evidence that the created BN model is consistent with Axiom 2. Additionally, it confirms the model's resilience.