Comparative Analysis of Danger Theory Variants in Measuring Risk Level for Text Spam Messages

Abstract

The issue of spam has been uprising since decades ago. Impact loss from various aspects has attacked the daily life most of us. Many approaches such as policy and guidelines establishment, rules and regulations enforcement, and even anti-spam tools installation appeared to be not enough to restrain the problem. To make things even worse, the spam’s recipients still easily get enticed and lured with the spam content. Hence, an advanced medium that acts as an implicit decision maker is desperately required to assist users to obstruct their eagerness responding against spam. The simulation of spam risk assessment in this paper is purposely to give some insights of how users can identify the imminent danger of received text spam. It is demonstrated by predicting the potential hazard with three different levels of risk (high, medium and low), according to its possible impact loss. A series of simulation has been conducted to visualize this concept using Danger Theory variants of Artificial Immune Systems (AIS), namely Dendritic Cell Algorithm (DCA) and Deterministic Dendritic Cell Algorithm (dDCA). The corpus of messages from UCI Machine Learning Repository has been deployed to illustrate the analysis. The outcome of these simulations verified that dDCA has consistently outperformed DCA in precisely assessing the risk level with severity concentration value for text spam messages. The findings of this work has demonstrated the feasibility of immune theory in risk measurement that eventually assisting users in their decision making.

Appendices

Appendix 1

Experiment 1. Characteristics of the testing setup for results tabulated in Table 5:

Table 5. Measurement of risk concentration for text spam messages using DCA and dDCA.

• Term weighting scheme: Term Frequency (TF).

• The value for risk scale (S1) and signal weight matrix (WM1) as tabulated in Table 1.

• For MCAV calculation, mature content referred to both High and Medium tokens.

• Text pre-processing is applied and antigen multiplication is not applied.

• See Sect. 5: Results and Analysis, Experiment 1; Table 2 and Fig. 3 for tabulated results and graph.

According to the risk concentration value tabulated in Table 5, it is empirically proven that DCA and dDCA are feasible algorithms that able to produce a risk level classification for text spam messages. The combination of more than one risky term may result in high risk such as containing URL that requested users to respond accordingly with the content of the message. Via this simulation, it is demonstrated that messages that contain information that requested users to respond (call, text, and chat) are in the appropriate high/medium risk category. While for messages that contain no requirement for users to respond (for instance a message with ID 15.txt) produced low-risk concentration level or potentially to be considered as a non-spam message.

Appendix 2

Experiment 4. Characteristics of the testing setup for results tabulated in Table 6:

Table 6. Measurement of risk concentration for text spam messages using immune classifier and non-immune classifier.

• The value for risk scale (S1) as tabulated in Table 1.

• Text pre-processing is applied.

• See Sect. 5: Results and Analysis, Experiment 4 and Fig. 6 for tabulated results and graph.

• Italic and underlined font indicates the falsely-classified for risk level

The value of confidence for the non-immune classifier is derived from the data mining tool, RapidMiner. This confidence value is referring to the probability of the message is tagged as spam. The risk level measured for the non-immune classifier is determined by mapping the confidence value of spam with the pre-set risk scale, S1.