An Intrusion Anomaly Detection Approach to Mitigate Sensor Attacks on Mechatronics Systems (original) (raw)

An Intrusion Anomaly Detection Approach to Mitigate Sensor Attacks on Mechatronics Systems

Wasswa Shafik
1{ }^{1} Computer Engineering Department Yazd University, Yazd, Iran
2{ }^{2} Digital Connectivity Research
Laboratory (DCRLab)
Kampala, Uganda
wasswashafik@ieee.org
S. Mojtaba Matinkhah*
Computer Engineering Department, Intelligence Connectivity Research Lab, Yazd University, Yazd, Iran matinkhab@yazd.ac.ir
Kassim Kalinaki
1{ }^{1} Department of Computer Science, Islamic University in Uganda
Mbale, Uganda
2{ }^{2} Digital Connectivity Research
Laboratory (DCRLab)
Kampala, Uganda
kalinaki@ieee.org

Abstract

Mechatronic systems (MES) have been widely studied and integrated into current smart engineering systems like robots, and control systems among others due to advance in technology. These systems are widely intruded on during operation through sensor attacks and their associated drawbacks. A robust technique for identifying and preventing sensor attacks in systems such as drones must be implemented in smart transportation networks. This paper proposes a novel intrusion anomaly detection approach (IADA) for MES sensors using recurrent neural networks. F1-score and One class classification (CM) anomaly detection was used to carry out performance assessment on several countermodels and classifiers. The results demonstrated that the proposed detection achieved 96%96 \% of F1-score, 99%99 \% of sensitivity, and 92%92 \% of precision in comparison to other counterparts across several drone platforms. The future research direction of the proposed model is also depicted.

Keywords- Deep Learning, Drone transportation, Intrusion Anomaly Detection, Mechatronics Systems, Sensor Attacks.

I. INTRODUCTION

Mechatronic systems are gaining more research interest from academia and the manufacturing industry due to the many peripherals that are attached to them to increase efficiency and productivity. These peripherals include actuators, controllers, vision components, and sensors, among others, as well as intelligent algorithms with learning capabilities running in the background. Mechatronics can either complement human operators or replace them in the performance of a growing number of activities. Therefore, decision-making, control and planning of modern MES encounter formidable obstacles [1]. Solutions like semi and adaptive learning methods have demonstrated their vigor in trajectory planning, recognition, perception, communication, situation comprehension and many more as a result of effective cross-disciplinary applications. This is because these approaches may adapt to their surroundings and gain knowledge from them.

The widespread use of underactuated mechatronic systems has become one of the most common ways to make things in the industry, where control effort and operation precision are both important parts of measuring performance. So, a big question for this research is how to control motion sensors in a way that stops sensor attacks and makes control easier at the same time. Even though energy optimization is possible with some open-loop algorithms, these solutions need linearization or approximation changes and are not very
robust [2]. Thus, the practical control performance of these systems is susceptible to decline.

Any of the machine’s moving elements could sustain structural damage. New techniques for monitoring their mechanical integrity would be an excellent way to approximate the structure’s strength, thus, its ability to effectively assist those in need or to improve the lives of the visually impaired. In addition, the alteration of a structure’s characteristics because of mechanical forces is irreversible [3].

Damage can cause, among other things, too much bending, buckling, and breaking and hence damage detection techniques give us information that we can act on right away, so we can fix things before a structure fails in a way that could be awful. If the affected area could be pinpointed, it would be helpful to act after getting a warning. This would help prevent structural collapse. This study shows how damage detection and localization techniques can be used to make a control system and an alert system that are both smart. The models explored were mechatronic systems used in healthcare.

Based on the current digital technology, algorithms for monitoring the mechatronic system of the rolling mill should be developed. In the periodic measurement mode, it would be advantageous to construct digital shadows of the system state parameters, also known as observers. The constant occurrence of unanticipated failures in rolling stand mechanical transmissions demonstrates the significance of this investigation. The angular gaps at the spindle joint enlarge as the mill is utilized for longer durations [4]. Consequently, it is a critical investigative trait that enables an evaluation of the transmission’s capacity to accomplish its intended purpose. Supervising these gaps in rolling stand MES is a substantial consideration in this study context.

As potential MES, upturned pendulum, response wheels and a helicopter with two degrees of freedom are both considered. To compensate for the problem that the reaction wheels have inverted the pendulum and are underactuated, a modified ultra-local prototype incorporating a reaction wheel velocity feedback has been developed. Such ultra-local limitations surrounding the volatile symmetry point are then determined using an algebraic estimator [5]. Adaptive MEScentered physiotherapy has played an essential role in enabling patients with gait disorders to walk independently

with no human being support. This is done to circumvent problems connected with standard medical therapy [6].

MES utilized in nuclear plants are regarded as safetyanalytical structures; therefore, precise assessments of the possibility of system failures are mandatory. Breakdowns of such systems rely on the proper operation of the component that comprises their foundation. The identification of component failure mechanisms is made possible by the development of models that can lead to the failure of systems. These types of failure scenarios are modelled as “fundamental events” in the system. To conduct precise measures of failure [7]. This refers to the circumstance in which one has the same level of knowledge about the malfunction parameters of otherwise indistinguishable mechanisms.

A common method of addressing intrusion anomaly detection (IAD) problems in the unmanned Aerial Vehicle domain is Long Short-Term Memory (LSTM) networkcentred anomaly detection (AD) [8-10]. But drone control systems vary widely, in relation to sixth-generation, sensors, and communication protocols [11] and [12]. Other notable studies related to this include [13-28]. Since drone sensor networks necessitate a variety of datasets, none of these approaches can effectively handle IAD [29] and [30]. Consequently, a technique that can amply consider these traits is required. Dissimilar existing AD, IAD arrive with the capability to learn from a training dataset with no anomalies in a prevalent way. Furthermore, the study employs an active drone flight logs of drone sensors, focusing on its on how it learns what is normal or not from the special drone to be controlled by the IADA. As a result, the suggested IADA method determines if messages are assaulted or known by examining typical drone sensor signal behaviors. This study contributes mainly as summarized below:

• Presents a novel intrusion anomaly detection approach (IADA) using recurrent neural networks for mechatronic systems, like drones.
• Uses unary classification or CM commonly referred to as one class classification to alleviate sensor attacks and associated concerns in drone systems with a variety of requirements.
• Validates the performance evaluation of numerous classifiers using simulated flight data of several drone models.

The remainder of this study is structured as follows. In Section 3, we present the proposed IADA method. In section 4 , the obtained simulated results and discussion are presented. Finally, section 5 has the conclusion and portrays our future research directions.

II. METHODOLOGY

To effectively solve this issue, this study makes use of the capabilities of artificial intelligence (AI) approaches, particularly autoencoders (AE), recurrent neural networks (ReNN), and unsupervised deep learning. Within this study, we considered ReNN reduced computation time and several parameters in comparison to the existing varieties. CM structure is an AI classification method that identifies data point anomalies from a certain imbalanced dataset and evaluates it with typical (recognized) class cases.

The approach further entails local outliers (LOF), AE, a support vector machine (OC-SVM), and Isolation Forest. Also, we consider that the known class in this instance consists of flight logs from the practice flight of drone sensor data utilized to carry out an operation. Any variant from the known is taken as an intrusion or fault anomaly. Within this paper, three approaches are considered, namely LOF, AE, and OC-SVM.

Fig. 1. Proposed Approach flow in detection and classification of intrusion anomaly signal

TABLE I. DRONES USED IN EXPERIMENTATIONS AND DATASET DESCRIPTION

TABLE II. COMPUTER SPECIFICATIONS AND SIMULATION ENVIRONMENT

The proposed method in this paper is to determine whether an incoming signal in a drone network is normal or not, dependent on the underlying IADs and classification system and is briefly demonstrated in Fig. 1 above. Per its distinct qualities, every classifier is treated to a particular analysis. The proposed model utilized five hidden layers to get the best optimum output. To train and test the network, we used the publicly available datasets making 0.7 of the data for training and 0.3 kept for testing. To improve the accuracy of the IADA model, we modified the last layer of the default ANN before testing the model.

The dataset contains abnormal (intrusion inclusion) (malicious) and normal (benign) reports with various attributes generated from an experimental setting consisting of six drone models, namely, 3DR Iris +, Yuneec H480, Standard Tailsitter, DeltaQuad Vertical Take-off and Landing (VTOL), Holybro S500 and standard plane using an open-source simulation tool PX4. The spoofing and jamming attacks are regarded as abnormal records. Also, feature clustering was used to get rid of useless features, using principal component analysis as detailed [5]. Table 1 presents a list of drones that have been used study and dataset description.

Table 2 presents the computer specification, simulation environment and tools. To simplify the confusion matrix (CoV)(\mathrm{CoV}), it is useful to properly envision methods using the four-ended table of True Negative (TN), False Negative (FN), True Positive (TP) and False Positive (FP) given that the following constraints are maintained to aid CoV to summarize the prediction.

• TP: The model correctly predicted an intrusion anomaly.
• TN: The model correctly predicted normal behavior.
• FP: The model incorrectly predicted an intrusion anomaly, when in fact it was normal behavior.
• FN: The model incorrectly predicted normal behavior, when in fact it was an intrusion anomaly.

The proposed model is tested on the AEs containing precision, sensitivity, F1- measurement, and the Threshold; these can be simply defined mathematically below:

A. Precision (Pr)

This is a positive predictive value, given by the number of true positives in the model argument in comparison to the number of positives that it asserts (1).

Pr⁡=TP⁡/(TP+FP)\operatorname{Pr}=\operatorname{TP} /(\mathrm{TP}+\mathrm{FP})

B. F1 Scores

The entire approach evaluation is calculated using this F1 measurement (3). It is computed as the average weight of the Pr⁡\operatorname{Pr} and the recall (2) of the model as expressed.

Recall =TP/(TP+FN) F1 Score =(2∗TP)/(2∗TP+FN+FP)\begin{aligned} & \text { Recall }=\mathrm{TP} /(\mathrm{TP}+\mathrm{FN}) \\ & \text { F1 Score }=\left(2 * \mathrm{TP}\right) /\left(2 * \mathrm{TP}+\mathrm{FN}+\mathrm{FP}\right) \end{aligned}

C. Sensitivity

One of the metrics that can be used to properly evaluate the categorization of a method is sensitivity which entails accuracy (4).

Sensitivity =(TN+TP)/(TP+FN+FP+TN)\text { Sensitivity }=(T N+T P) /(T P+F N+F P+T N)

III. RESULTS AND DISCUSSIONS

This section entails the simulated results and the performance evaluation of the considered counterparts of the study. The proposed model evaluation based on the CM-AE model employing F1, sensitivity, and precision is displayed across the six drone models in Table 3’s results. Every model’s behavior is evaluated according to its F1-scores variables.

TABLE III. INTRUSKIN ANOMALY DETECTIONS ACCORDING TO THE AUTOENCODER (AVG; AVERAGE, FOR SENSITIVITY, PRECISION AND THRESHOLD ARE COMPUTED IN PERCENTAGES)

TABLE IV. INTRUSION ANOMALY DETECTIONS ACCORDING TO THE MODELS (ONE CLASS CLASSIFICATION-BASED SUPPORT VICTOR MACHINE, AUTOENCODER, AND LOCAL OUTLIERS)

The AE successfully distinguished between a genuine sensor reading and a faked or jammed reading with an average precision, sensitivity, and F1-score value of 92%92 \%, 99%99 \%, and 96%96 \%, respectively. This confirms that the AE is capable of adequately identifying and thwarting drone sensor attacks. This implies that the sensor intrusion can be detected more effectively using advanced classifiers.

Fig. 2. Sensor Attack Detections according to used classifiers (One class classification-based support vector machine, autoencoder, and local outliers)

The average F1-Scores are demonstrated in Fig. 2 for the sensor attack detection. It is demonstrated that the OC-SVM, OC-AE, and LOF attained 82%,96%82 \%, 96 \%, and 59%59 \%, respectively. The results attest that the autoencoder fundamentally performs better as compared to other detection approaches in drone IADA. Jammed or spoofed sensor signals on numerous drone networks can be significantly spotted and prevented.

The overall performance is briefly presented in Table 4. Also, shows that the average F1-score for LOF is 59%59 \%, for OC-SVM it is 82%82 \%, and for OC-AE it is 96%96 \% with a sensitivity of 0.99%0.99 \%. These findings demonstrate that a faked or jammed sensor signal on various drone networks may be
significantly detected and avoided when a dehazing OC-AE is used as the underlying classification algorithm in a drone IADA.

IV. CONCLUSION

This paper proposed IADA that uses ReNN to identify sensor threats on drone networks across various configurations. Based on the comprehensive technique for preventing and detecting attacks on mechatronics Systems, this research arena is gaining much momentum in academia and industrialization. The dataset consisted of six drone models’ regular and malicious (spoofing and jamming) flight logs. The acquired result(s) demonstrated in third and fourth Tables showed that AE can satisfactorily detect intrusions to the drone sensor signals. For our future research based on this paper, the proposed model is to be tested for onboard drone sensor configurations. Also, the model is to be tested on the rest of the two anomaly categories, namely, collective, and contextual (context) anomalies besides the tested point (data point) anomaly. In future, we compare the model with more parameters in addition to LoF, AE, OC-SVM, and other types that are still open for the model to test like logistic regression, decision trees, and Random Forest, among others. Lastly, on model evaluation, other parameters are still available for testing, like the receiver operating traits curve.
Author Contributions: Conceptualization, W.S., S. M. M., K.K; Data Acquisition, W.S., K.K; Methodology, K.K and W. S.; S. M. M. Supervision, W.S; Writing-original draft, W.S. K.K.; Writing-review and editing, W.S. K.K. All authors have read and agreed to the published version of the manuscript.

DataAvailability:https://github.com/wasswashafik/Intrusion -Anomaly-Detection-Approach
Funding: This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Conflicts of Interest: No declared conflicts of interest regarding this paper.

REFERENCES

[1] H. Gao, B. Liang, R. Oboe, Y. Shi, S. Wang and M. Tomizuka, “Guest editorial introduction to the focused section on adaptive learning and control for advanced mechatronics systems,” IEEE/ASME Transactions on Mechatronics, vol. 27, no .2, pp. 607610, 2022, https://doi.org/10.1109/TMECH.2022.3151663.
[2] T. Yang, N. Sun, H. Chen and Y. Fang, "Adaptive optimal motion control of uncertain underactuated mechatronic systems with actuator constraints. IEEE/ASME Transactions on Mechatronics, pp. 1 - 13, 2022, https://doi.org/10.1109/TMECH.2022.3192002.
[3] C. Dragne, V. Chiroiu, M. Iliescu and I. Todirite, “Damage detection and smart warning for eventual structure failures in mechatronic systems,” IEEE World Conference on Applied Intelligence and Computing (AIC), Sonbhadra, India, pp. 467-472, 2022, IEEE, https://doi.org/10.1109/AIC55036.2022.9848855.
[4] O. A. Gasiyarova, A. S. Karandare, I. N. Erdakov, B. M. Loginov and V. R. Khranshin, “Developing digital observer of angular gaps in a rolling stand mechatronic system,” Machines, vol. 10, no. 2, pp. 141, 2022, https://doi.org/10.3390/machines10020141.
[5] G. Pereira das Neves and B. Augusto Angélico, “Model - free control of mechatronic systems based on algebraic estimation,” Asian Journal of Control, vol. 24, no. 4, pp. 1575-1584, 2022, https://doi.org/10.1002/asjc.2596.
[6] J. P. A. Joel, R. J. S., Raj and N. Muthukumaran, “Review on gait rehabilitation training using human adaptive mechatronics system in biomedical engineering,” International Conference on Computer Communication and Informatics (ICCCI), Hammamet, Tunisia, pp. 15, 2022, IEEE, https://doi.org/10.1109/ICCCI54379.2022.9740794.
[7] P. Singh, and L. K. Singh, “State of knowledge correlation in failure analysis of mechatronics systems,” IEEE Transactions on Reliability, 2022, https://doi.org/10.1109/TR.2022.3172565.
[8] S. Mehtab and J. Sen, “Analysis and forecasting of financial time series using CNN and LSTM-based deep learning models,” Advances in Distributed Computing and Machine Learning, pp. 405-423, Springer, Singapore, 2022, https://doi.org/10.1007/978-981-16-48076\_39.
[9] Y. N. Zhou, S. Wang, T. Wu, L. Feng, W. Wu et al., “For-backward LSTM-based missing data reconstruction for time-series Landsat images,” GIScience & Remote Sensing, vol. 59, no. 1, pp. 410-30, https://doi.org/10.1080/15481603.2022.2031549.
[10] W. Shafik, “Cyber Security Perspectives in Public Spaces: Drene Case Study,” In Handbook of Research on Cybersecurity Risk in Contemporary Business Systems, pp. 79-97, 2023. IGI Global.
[11] B. Fraser, S. Al-Rubaye, S. Aslam and A. Tsourdos, “Enhancing the security of unmanned aerial systems using digital-twin technology and intrusion detection,” IEEE/AIAA 40th Digital Avionics Systems Conference (DASC), pp. 1-10, San Antonio, TX, USA, 2021, IEEE. https://doi.org/10.1109/DASC52595.2021.9594321.
[12] Y. Han, G. Song, F. Liu, Z. Geng, B. Ma et al., “Fault monitoring using novel adaptive kernel principal component analysis integrating grey relational analysis,” Process Safety and Environmental Protection, vol. 157, 2022, pp. 397-410, 2022. https://doi.org/10.1016/j.psep.2021.11.029.
[13] W. Jiang, “A machine vision anomaly detection system to industry 4.0 based on variational fuzzy autoencoder,” Computational Intelligence and Neuroscience, 2022, https://doi.org/10.1155/2022/1945507.
[14] C. Dragne, V. Chiroiu, M. Iliescu, and I. Todirite, “Damage detection and smart warning for eventual structure failures in mechatronic systems,” in 2022 IEEE World Conference on Applied Intelligence and Computing (AIC), Sonbhadra, India, 2022, pp. 467-472, https://doi.org/10.1109/AIC55036.2022.9848855.
[15] F. Mhenni, F. Vitolo, A. Rega, R. Plateaux, P. Hehenberger et al., “Heterogeneous models integration for safety critical mechatronic
systems and related digital twin definition: application to a collaborative workplace for aircraft assembly,” Applied Sciences, vol. 12, no. 6, pp. 2787, 2022, https://doi.org/10.3390/appl2062787.
[16] C. Dragne, I. Todirite, M. Iliescu, and M. Pandelea, “Distance assessment by object detection-for visually impaired assistive mechatronic system,” Applied Sciences, vol. 12, no. 13, pp. 6342, 2022, https://doi.org/10.3390/appl12136342.
[17] Y. Jun, A. Craig, W. Shafik and L. Sharif, “Artificial intelligence application in cybersecurity and cyberdefense,” Wireless Communications and Mobile Computing, vol. 2021. https://doi.org/10.1155/2021/3329581.
[18] W. Shafik, S. M. Matinkhah, M. N. Sanda, and F. Shokoor, “Internet of things-based energy efficiency optimization model in fog smart cities,” JOIF: International Journal on Informatics Visualization, vol. 5, no. 2, pp. 105-112, 2021.
[19] L. Qian, Q. Pan, Y. Lv, and X. Zhao, “Fault detection of bearing by resnet classifier with model-based data augmentation,” Machines, vol. 10, no. 7, pp. 521, 2022, https://doi.org/10.3390/machines10070521.
[20] W. Shafik, S. M. Matinkhah, and M. Ghasemzadeh, “Internet of things-based energy management, challenges, and solutions in smart cities,” Journal of Communications Technology, Electronics and Computer Science, vol. 27, pp. 1-11, 2020. http://dx.doi.org/10.22385/jctecs.v27i0.302.
[21] M. Yu, D. Lan, C. Jiang, B. Xu, D. Wang, and R. Zhu, “Hybrid condition monitoring of nonlinear mechatronic system using biogeography-based optimization particle filter and optimized extreme learning machine,” ISA transactions, vol. 120, pp. 342-359, 2022, https://doi.org/10.1016/j.isatra.2021.03.018.
[22] W. Shafik, S. M. Matinkhah, and M. Ghasemzadeh, “Theoretical understanding of deep learning in uav biomedical engineering technologies analysis,” SN Computer Science, vol. 1, no. 6, pp. 1-13, 2020. https://doi.org/10.1007/s42979-020-00323-8.
[23] F. Shokoor, W. Shafik, and S. M. Matinkhah, “Overview of 5G & beyond security,” EAI Endorsed Transactions on Internet of Things, vol. 8, no. 30, 2022.
[24] W. Shafik, M. Matinkhah, and M. N. Sanda, “Network resource management drives machine learning: a survey and future research direction,” Journal of Communications Technology, Electronics and Computer Science, pp. 1-5, 2020.
[25] Y. Furukawa and M. Deng, “SVM-based fault detection for double layered tank system by considering ChangeFinder’s characteristics,” International Journal of Advanced Mechatronic Systems, vol. 9, no. 4, pp. 185-192, 2022, https://doi.org/10.1504/IJAMECHS.2022.123141.
[26] S.N. Alaziz, B. Albayati, A. A. El-Bagoury, and W. Shafik, “Clustering of COVID-19 Multi-Time Series-Based K-Means and PCA With Forecasting,” International Journal of Data Warehousing and Mining (IJDWM), vol. 19, no. 3, pp. 1-25.
[27] W. Shafik, S. M. Matinkhah, S. S. Afolabi, M. N. Sanda, "A 3dimensional fast machine learning algorithm for mobile unmanned aerial vehicle base stations, " International Journal of Advances in Applied Sciences, vol. 2252, 2020.
[28] W. Shafik, and S. M. Matinkhah, "Admitting new requests in fog networks according to erlang b distribution, " In2019 27th Iranian Conference on Electrical Engineering (ICEE), 2019. IEEE.
[29] I. Ilie, M. N. Ardeleanu, and V. A. Soare, “Mems measurement using a plastic and metal arm integrated into a mechatronic system,” Int. J. Mechatron. Appl. Mech, vol. 11, pp. 259-265, 2022.
[30] S. Wang, B. Yang, H. Chen, W. Fang and T. Yu, “LSTM-based deformation prediction model of the embankment dam of the danjiangkou hydropower station,” Water, no. 14, no. 16, 2464. https://doi.org/10.3390/w14162464.