2024 5th IEEE Global Conference for Advancement in Technology (GCAT)

Karnataka, India. Oct 4-6, 2024

Detecting Hardware Trojans in IoT Devices Using

Semi-Supervised Learning with XG-Boost Classifier.
Hrithik Raj N Nirmala Devi M Vaishnavi sankar
Department of Electronics and Department of Electronics and Department of Electronics and
Communication Engineering. Communication Engineering. Communication Engineering.
PES University PES University PES University
Bangalore, India Bangalore, India Bangalore, India
email : [email protected] email : [email protected] email : [email protected]

Abstract— The increasing rate of hardware Trojans poses abnormalities that may indicate the presence of hardware
critical security risks to chips and integrated circuits (ICs) used in Trojans by lowering the dimensionality of the data. A detection
high precedence applications. Conventional detection methods accuracy rate of greater than 95% is what we aim to achieve. We
often fall short due to the growing complexity of Trojans and the verify our method and demonstrate its great precision and
unfeasibility of extensive testing. This study introduces a adaptability with comprehensive tests on various chips. The
comprehensive approach to extract key features from netlist data, outcomes demonstrate the scalability and efficacy of our
such as connectivity, levels, fan-in, fan-out, and counts of flip- approach in strengthening IC security in the face of hardware
flops/multiplexers, essential for identifying hardware Trojans. Trojans. The classifier's capacity to discriminate between benign
Principal Component Analysis (PCA) was applied to reduce
and malevolent changes is enhanced by this integration.
feature dimensions while preserving 95% variance, and
permutation importance was utilized to determine feature A. Existing Methods and their Problems.
importance. Furthermore, Label Spreading was employed to
make use of unlabeled data, improving the training dataset and
Extensive taxonomy is given in [2], [3]. The scoring-based
model performance. To ensure optimal selection of approach is proposed in [8], where a threshold is set and scores
hyperparameters and accurate net predictions, Grid-Search CV are assigned to all nine weak nets for the purpose of
was used to fine-tune an XG-Boost classifier. By combining classification. Because this approach is threshold dependent,
supervised and unsupervised learning, our method effectively sophisticated Trojans are missed.
detects differences that indicate the presence of hardware Trojans The machine learning (ML) method is put forth by [4]. It
with minimal labeled data. The effectiveness and viability of semi- extracts five structural features from the netlist and applies
supervised hardware Trojan detection were highlighted by the
Support Vector Machine (SVM) to achieve detection. However,
suggested approach's average true positive rate of 91.7% and true
negative rate of 94.77% when evaluated on the Trust-HUB
the class weights are inefficient, resulting in notable false
benchmarks. positives and negatives across several benchmarks. Additionally,
they have extracted 51 features from [5] in order to improve
Index Terms— Hardware Trojans, Integrated Circuits (ICs), detection, and they use random forest (RF) to select the top 11
Feature Extraction, Principal Component Analysis (PCA), Label features for detection. However, because of its heavy weighting
Spreading, Semi-supervised Learning, XG-Boost Classifier, Grid- of fewer predictive markers, the RF also fails to identify Trojans
Search CV, Trust-HUB Benchmarks, Anomaly Detection in some benchmarks.

I. INTRODUCTION Additionally, by repeatedly using the Trojan samples, they

employ oversampling to balance the classes, which increases the
Hardware Malicious alterations known as Trojans have the likelihood of overfitting in the model. There have been attempts
potential to compromise the security and operation of integrated at an unsupervised HTD technique known as PL-HTD, which
circuits (ICs), which are utilized in several applications. uses a local outlier factor algorithm to generate the best feature
Although side-channel analysis and comprehensive testing are set for principal component analysis-based unsupervised
useful approaches for identifying Trojans, their increasing classification. [13]. Additionally, a Neural Network (NN) based
complexity calls for the creation of more adaptable and method [9] is also suggested; it employs the same kind of
trustworthy detection technologies. weighting strategy as in [4] and uses the features of [5] for
Machine learning (ML) has been a huge help with these detection, resulting in lower true positive rates. In [10] and [11],
problems, especially with semi-supervised learning that uses testability metrics and transition probability were investigated in
both labeled and unlabeled data. This approach is particularly more detail. Low TPR was the outcome of limited representative
effective for hardware Trojan identification because labeled data training data. Testability measure-based characteristics and
is hard to come by. In this study, we propose a novel detection structural features were added by Liu et al. [12], [13] for better
method that combines the XG-Boost classifier with semi- Trojan identification. The method is computationally intensive,
supervised learning. The well-known data classification and its temporal complexity grows with circuit size. To solve the
algorithm XG-Boost is used to identify patterns associated with problem of feature correlation and data imbalance for XG-Boost,
certain hardware components. To enhance feature extraction, we a feature selection and class weighting technique was developed
employ the PI-PCA (Principal Component Analysis) tool in our [20]. An unsupervised HTD technique called PL-HTD has been
workflow. PI-PCA can identify significant traits and attempted; it uses a local outlier factor algorithm to generate the

979-8-3503-7668-5/24/$31.00 ©2024 IEEE 1

optimal feature set for principal component analysis-based better by utilizing a considerable quantity of unlabeled input.
unsupervised classification [19]. Due to the method's limited This methodology improves design optimization and reliability
ability to generalize, there were a lot of false positives. assessment by providing an efficient and scalable means of
analyzing integrated circuit designs.
Trojan circuit triggering characteristics and a feature
analysis method based on a flip-flop level information flow C. Contributions of this Paper
graph are described in [14] and [15]. In [16], an attempt was The contributions of this paper are summarized as follows:
made to detect hardware Trojans using few-shot learning. Poor
generalization capacity occurs from the findings not matching • A thorough method for identifying important aspects in
published results in terms of true positive rate (TPR), even netlist data extraction is provided. In order to detect
though the objective is to construct an HTD based on a similarity hardware Trojans, these features which include
function. Supervised algorithms are only fully functional in the connection, levels, fan-in, fan-out, and counts of flip-flops
presence of high-quality labeled data. and multipliers are crucial.
In [17] and [18], an attempt to integrate static and dynamic • The data's dimensionality was decreased while 95% of the
features was made. Because of the limited Trojan space that can variance was preserved with the application of Principal
be learned during the training phase, techniques that use Component Analysis (PCA). Each attribute's importance
structural features perform worse in terms of true positive rate was also assessed using permutation importance.
(TPR), which leads to poor generalization capacity, even though • The model makes use of labeled and unlabeled data to
the method had produced 95% average TPR in Trojan detection. improve its predictive power by leveraging the benefits of
Supervised algorithms are only fully functional in the presence semi-supervised learning. The model can adapt to different
of high-quality labeled data. Labeling every piece of circuit data types of data and offers a comprehensive solution for
by hand gets tedious and time-consuming. hardware Trojan detection when Label Spreading and XG-
The problem is made worse by the increasing complexity of Boost are combined.
circuits. On the other hand, unsupervised algorithms cannot • The model's accuracy and high efficiency are guaranteed
accurately infer patterns that disclose Trojan traits without vast by the integration of Grid-Search CV, which qualifies it for
amounts of data. Consequently, a system that circumvents the real-time applications aimed at identifying Trojans in
drawbacks of both approaches becomes essential when Internet of Things devices.
considering the range of threat scenarios. • Grid-Search CV was used to optimize the hyperparameter
B. Summary of the Proposed Method selection for an XG-Boost classifier and to deliver a
dependable net criticality prediction.
The recommended method concentrates on identifying
significant traits in netlist data to provide a reliable dataset for II. METHODOLOGY FOR HARDWARE TROJAN DETECTION
hardware Trojan identification. translating Verilog to netlist, USING SEMI-SUPERVISED LEARNING WITH XG-BOOST
obtaining relevant features from an Excel file, configuring net CLASSIFIER
properties, and translating netlist data back into Verilog are the
initial phases in the process. Numerous computed attributes are A. Definition of Trojan Features in a Netlist
present in each net, such as multiplexer connections, flip-flop, Hardware Trojans (HTs) can only be found by extracting a
connectivity, level, and fan-in and fan-out. Power metrics are certain set of structural attributes from the gate-level netlists of
handled by processes that depend on specific models to do the input circuits that were seized. The number of features we
calculations correctly. Following collection, the features are select is important when it comes to detecting Trojans. Once a
organized into a Data-Frame, standardized using Standard segment of the circuit has been identified, its underlying
Scaler, and then PCA is applied to minimize the dimensionality structure remains constant, producing consistent feature values.
while preserving 95% of the variance. But if a Trojan infection spreads to this sector, some of these
A classifier is developed using XG-Boost in a semi- numbers will show notable deviations. This section introduces
supervised learning approach that uses Label Spreading to five aspects associated with Trojans:
leverage unlabeled data for classification. Using this process, the • Level: The distance from primary input.
dataset is split into labeled and unlabeled sets, with 80% of the
dataset being categorized as unlabeled. By distributing labels • Connectivity: Number of gates connected to the net.
from identifiable data to unlabeled scenarios, Label Spreading
• Primary input: Level the net is from the primary input.
enhances the training dataset. The XG-Boost classifier is then
evaluated and trained on the combined dataset, and its • Primary output: Level the net is from the primary output.
hyperparameters are adjusted using Grid-Search CV. Many
performance indicators, such as accuracy, precision-recall curve, • Fan-in_X: The distance from primary input.
ROC curve, and confusion matrix, are used to evaluate the • In/Out flip-flop_X: The number of flipflops connected
model. X level from input/output side of net.
The outcomes validate our feature extraction method's • In/out multiplexer: The number of muxes connected X
effectiveness and the classification model's dependability in level from input/output side of net.
forecasting integrated circuit (IC) net criticality. When semi-
supervised learning techniques are applied, the model performs

2
B. Proposed Methodology. other pertinent parameters including level, connectivity, primary
A well-defined set of steps, including netlist construction, inputs, and primary outputs, during the data preparation process.
data preparation, feature extraction, dimensionality reduction, Next, the dataset is divided into characteristics and target
semi-supervised learning, classification, and model evaluation, variables. To guarantee uniform scaling, feature values are
make up the suggested methodology for locating hardware normalized.
Trojans in integrated circuit (IC) designs. The first step in the As part of the model evaluation procedure, predictions are
method is to extract the circuit's netlist by utilizing the Cadence produced and projected probabilities for the test set are
Genus tool to translate Verilog code into netlist data. This netlist calculated using the upgraded classifier. The measures used to
lays the foundation for further investigation by giving a evaluate the model's performance include accuracy, ROC curve,
thorough explanation of every component of the circuit and how Precision-Recall curve, confusion matrix, and a classification
it is connected. The dataset is then loaded from an Excel file report. These tools allow an in-depth understanding of the
holding the netlist data of integrated circuit designs, along with model's effectiveness and dependability in finding hardware
other pertinent parameters including level, connectivity, primary Trojans. This methodology offers a structured approach to
inputs, and primary outputs, during the data preparation process. collecting, analyzing, and categorizing IC design data in order
Next, the dataset is divided into characteristics and target to detect hardware Trojans. Through the use of semi-supervised
variables. To guarantee uniform scaling, feature values are learning for label propagation and PCA for dimensionality
normalized. The process of finding and acquiring the primary reduction, along with thorough model validation, this
inputs and outputs is known as feature extraction, and it comes methodology guarantees great performance and security in IC
after that. design analysis.
Each net is initialized with attributes such as connectivity,
III. HARDWARE TROJAN CLASSIFICATION RESULT
fan-in/out counts, and placeholders for power values.
A. Effective Feature Extraction and Data Preparation
PCA reduces the dataset's dimensionality while maintaining
its most significant features, ensuring that the classification
model focuses on the most crucial data. By taking this step, the
model's performance is enhanced and its computational load is
reduced. Insufficient data points are eliminated in order to
manage missing values in the target variable. Standard-Scaler is
used to standardize features in order to guarantee that they have
zero mean and unit variance, which is a prerequisite for machine
learning techniques, particularly XG-Boost.
B. Robust Classification Using XG-Boost
Optimizing hyperparameters and achieving balanced class
weighting XG-Boost is started with balanced class weights in
order to address class imbalance and guarantee that majority and
minority classes are treated equally. By modifying
hyperparameters such the number of estimators, maximum
depth, learning rate, subsample ratio, and column sampling by
tree, you may use GridSearchCV to optimize the model for best
performance. The model's effectiveness is evaluated using a
variety of metrics, such as recall, accuracy, precision, F1 score,
and True Positive and True Negative rates (TPR, TNR, FPR, and
FNR). Plotting the ROC curves and associated areas under the
curves for each class allows one to see the discriminative power
of the model.
Fig. 1. Flow of the Proposed HT Detection and Diagnosis Method. C. Detailed Performance Analysis
Sk-learn produces a confusion matrix that shows the counts
A well-defined set of steps, including netlist construction, of true positives (TP), false positives (FP), true negatives (TN),
data preparation, feature extraction, dimensionality reduction, and false negatives (FN). A better picture of the model's
semi-supervised learning, classification, and model evaluation, performance in terms of prediction accuracy for each class is
make up the suggested methodology for locating hardware given by normalizing the matrix to display the percentages of
Trojans in integrated circuit (IC) designs. The first step in the each category. Using roc-curve from Sk-learn, ROC curves for
method is to extract the circuit's netlist by utilizing the Cadence classes 0 and 1 are plotted. The trade-off between the genuine
Genus tool to translate Verilog code into netlist data. This netlist positive rate and the other is shown visually by the ROC
lays the foundation for further investigation by giving a (Receiver Operating Characteristic) curve. (TPR) as well as the
thorough explanation of every component of the circuit and how false positive rate (FPR) for a binary classifier when the
it is connected. The dataset is then loaded from an Excel file discrimination threshold makes changes. An evaluation of the
holding the netlist data of integrated circuit designs, along with model's performance across all categorization thresholds is

3
facilitated by the ROC curve? The classifier's performance in
identifying Class 0 (Negative) and Class 1 (Positive) cases is
represented by two curves depicted in the given ROC curve.
Shown in blue is the ROC curve for Class 0. This class has a
0.93 AUC (Area Under the Curve). Classifier performance in
differentiating between Class 0 and Class 1 cases is good, as
evidenced by its AUC of 0.93. Accordingly, the classifier
minimizes the amount of positive occurrences it wrongly labels
as negative while being extremely successful at accurately
recognizing negative examples.
The shape of the ROC curve for Class 0 demonstrates that
the classifier achieves a high TPR while maintaining a low FPR.
The curve rises sharply towards the top-left corner of the plot,
which is characteristic of a strong classifier. This sharp rise
indicates that the classifier quickly achieves high sensitivity
(true positive rate) with only a small increase in the false positive
rate. The ROC curve for Class 1, shown in orange, also has an
Fig. 3. ROC curve of RS232-T1100
AUC of 0.93. Similar to Class 0, this high AUC value signifies
that the classifier is proficient at distinguishing Class 1 instances
strengths and weaknesses. The model's strong capacity to
from Class 0 instances. The classifier effectively identifies
reliably identify instances of hardware Trojans is demonstrated
positive instances, while keeping the rate of incorrectly labeled
by its True Positive (TP) rate of 96.97%. The majority of real
negative instances low. The performance of the classifier for
Trojan instances are recognized and correctly reported by the
Class 1 is equally strong as it is for Class 0.
model thanks to this high TP rate, which is essential for
Class 1's curve likewise climbs sharply in the direction of the preserving the security of IoT devices. The model occasionally
upper-left corner, showing a high true positive rate and a low sets up false alarms by mistakenly labeling safe circuits as
false positive rate. This sharp increase demonstrates that the harboring Trojans, as seen by the 3.03% False Positive (FP) rate.
classifier can identify positive examples with high accuracy and In order to increase the model's efficiency and trustworthiness in
little false positive trade-off. Class 0 and Class 1's high AUC real-world settings, a low false positive rate is crucial for
values show how well the classifier performs overall in reducing pointless inquiries and interventions. The model
differentiating between the two classes. A thorough misses a large percentage of real Trojan occurrences, as
understanding of the classifier's effectiveness is provided by the evidenced by the 13.04% False Negative (FN) rate. This flaw
ROC curve, which demonstrates that it can correctly recognize suggests that the model is missing some harmful circuitry
both positive and negative examples. A robust and dependable alterations, which might be dangerous if Trojans are used
hardware model is indicated by the high AUC values and the without detection. It is critical to lower the FN rate in order to
steep rise of the curves towards the top-left corner of the figure. guarantee thorough identification and mitigation of potential
Trojan detection. security issues. The model's overall dependability in properly
recognizing non-Trojan cases is shown in the True Negative (TN)
The XG-Boost classifier's effectiveness in identifying rate of 86.96%. Although this shows a reasonably high degree
hardware Trojans in RS232-T1000 circuits is comprehensively of accuracy in identifying benign circuits, further optimization
summarized in the confusion matrix. The confusion matrix's may be done to improve detection accuracy and lower the rates
evaluation criteria play a critical role in identifying the model's of both false positives and false negatives.
The XG-Boost classifier's merits are shown by the confusion
matrix, especially its low false alarm rate and high TP rate,
which highlight how well it detects hardware Trojans with little
false positives. However, the FN rate indicates that more work
has to be done to hone the model's sensitivity and guarantee that
every Trojan occurrence is appropriately identified. The model's
dependability is confirmed by the TN rate, although improving
the model's overall detection accuracy is still a major goal. Using
the RS232-T1100 dataset, we trained an XG-Boost classifier
with characteristics extracted from an Excel source. Grid Search
Cross-Validation was utilized for optimization to increase the
classifier's capacity for prediction. Our examination covered a
wide range of factors that are crucial for assessing categorization
performance. Specifically, the model demonstrated 93.20%
accuracy on the test dataset, suggesting that it can detect
instances with precision.

Fig. 2. Confusion matrix for RS232-T1100.

4
 The associated classification report includes Precision, Score of 90.20%, the model demonstrated good performance in
Recall, F1-score, and Support metrics for both Positive and balancing recall and precision metrics. The classifier's ability to
Negative classes, giving a comprehensive knowledge of the discriminate between Positive and Negative classes was proved
classifier's precision-recall trade-offs and anticipated by visible analysis that employed Receiver Operating
dependability. The proportion of accurate and inaccurate Characteristic (ROC) curves and associated Area Under Curve
classifications across classes was displayed using a normalized (AUC) values. This thorough evaluation method offers an
confusion matrix, which further facilitated a detailed assessment unbiased analysis of the XG-Boost classifier's performance on
of the model's robustness. Key performance indicators were the RS232-T1600 dataset, which is crucial for applications in the
derived to assess the classifier's sensitivity and specificity, technology and engineering sectors that require precise
respectively. Among these measures were the True Negative predictive modeling.
Rate (94.70%) and the True Positive Rate (92.47%). With an F1
TABLE I. RESULTS OF XG-BOOST ALGORITHM AND PROPOSED TECHNIQUE ON TRUST-HUB BENCHMARKS (%)

TRUST-HUB BENCHMARKS ACCURACY PRECISION RECALL f-MEASURE TPR TNR

RS232-T1000 98.2 98 97.3 98 97.6 100

RS232-T1100 93 93.2 90 93.1 90 96.6
RS232-T1200 91 91 93.8 92 98.3 88.8
RS232-T1300 94.6 95.6 94 95 94.1 96
RS232-T1400 91 91.2 90.3 88 90.7 91
RS232-T1500 93.2 93 87 91.2 88 96.8
RS232-T1600 91 91.2 88.6 90.2 88.6 94
Average 93.1 93.3 91.5 92.5 92.4 94.7

TABLE II. AVERAGE COMPARATIVE RESULTS OF DIFFERENT

EVALUATION MEASURES ON TRUST-HUB BENCHMARKS (%)

provides insights useful in situations when precise classification

EXISTING MODEL is needed in the actual world.
EVALUATIO PROPOS
N ED REFERENCES
MEASURES TECHNI
QUE [1] V. Sankar, M. Nirmala Devi. and M. Jayakumar., "Data Augmented Hardware
[4] [5] [7] [9] [14] [17] Trojan Detection Using Label Spreading Algorithm Based Transductive
Learning for Edge Computing-Assisted IoT Devices," in IEEE Access, vol. 10,
pp. 102789-102803, 2022, doi: 10.1109/ACCESS.2022.3209705.
Accuracy 86.9 89.7 90.6 80 96.1 96.5 93.1
[2] H. Li, Q. Liu, and J. Zhang, “A survey of hardware Trojan threat and defense,”
Integration, vol. 55, pp. 426–437, 2016.
Precision 83 88 92 75 81.4 84.5 98.86 [3] Z. Huang, Q. Wang, Y. Chen, and X. Jiang, “A survey on machine learning
against hardware Trojan attacks: Recent advances and challenges,” IEEE
Recall 90 90.7 86.6 80 81.2 82.5 99 Access, vol. 8, pp. 10 796–10 826, 2020.
[4] K. Hasegawa, M. Oya, M. Yanagisawa, and N. Togawa, “Hardware Trojans
F-measure 86 89.3 89.2 77.4 80 82.7 98.76 classification for gate-level netlists based on machine learning,” in On-Line
Testing and Robust System Design (IOLTS), 2016 IEEE 22nd International
Symposium on. IEEE, 2016, pp. 203–206.
IV. CONCLUSION
[5] K. Hasegawa, M. Yanagisawa, and N. Togawa, “Trojan-feature extraction at
Using an XG-Boost classifier, we examined the RS232- gate-level netlists and its application to hardware-Trojan detection using
T1600 dataset with an emphasis on technological and random forest classifier,” in Circuits and Systems (ISCAS), 2017 IEEE
engineering applications. We used a method that involved International Symposium on. IEEE, 2017, pp. 1–4.
careful preparation of the data, deliberate division of the dataset, [6] L. Kampel, P. Kitsos, and D. E. Simos, ‘‘Locating hardware Trojans using
and extensive hyperparameter adjustment using Grid Search combinatorial testing for cryptographic circuits,’’ IEEE Access, vol. 10, pp.
Cross-Validation. The classifier's resilience in class label 18787–18806, 2022.
prediction was demonstrated by its remarkable 93.20% accuracy [7] Nigh and A. Orailoglu, ‘‘AdaTrust: Combinational hardware Trojan
on the test set. Its efficacy in differentiating between classes was detection through adaptive test pattern construction,’’ IEEE Trans. Very
Large Scale Integr. (VLSI) Syst., vol. 29, no. 3, pp. 544–557, Feb. 2021.
shown by precise measurements like Precision, Recall, and
[8] M. Oya, Y. Shi, M. Yanagisawa, and N. Togawa, “A score-based classification
targeted performance indicators like True Positive Rate (92.47%) method for identifying hardware-Trojans at gate-level netlists,” in Proceedings
and True Negative Rate (94.70%). This study adds a solid of the 2015 Design, Automation & Test in Europe Conference & Exhibition.
framework for using XG-Boost in predictive modeling and EDA Consortium, 2015, pp. 465–470.

5
[9] K. Hasegawaa, M. Yanagisawa, and N. Togawa, “Hardware Trojans
classification for gate-level netlists using multi-layer neural networks,” in On-
Line Testing and Robust System Design (IOLTS), 2017 IEEE 23rd
International Symposium on. IEEE, 2017, pp. 227–232.
[10] Y. Su, H. Shen, R. Lu, and Y. Ye, ‘‘A stealthy hardware Trojan design and
corresponding detection method,’’ in Proc. IEEE Int. Symp. Circuits Syst.
(ISCAS), May 2021, pp. 1–6.
[11] K. Huang and Y. He, ‘‘Trigger identification using difference-amplified
controllability and dynamic transition probability for hardware Trojan
detection,’’ IEEE Trans. Inf. Forensics Security, vol. 15, pp. 3387–3400,
2020.
[12] Q. Liu, P. Zhao, and F. Chen, ‘‘A hardware Trojan detection method based
on structural features of Trojan and host circuits,’’ IEEE Access, vol. 7,
pp. 44632–44644, 2019.

[13] Dong, Y. Liu, J. Chen, X. Liu, W. Guo, and Y. Chen, ‘‘An unsupervised
detection approach for hardware Trojans,’’ IEEE Access, vol. 8, pp. 158169–
158183, 2020.
[14] K. Hasegawa, M. Yanagisawa, and N. Togawa, ‘‘A hardware-Trojan classi-
fication method utilizing boundary net structures,’’ in Proc. IEEE Int. Conf.
Consum. Electron. (ICCE), Jan. 2018, pp. 1–4.
[15] T. Kurihara and N. Togawa, ‘‘Hardware-Trojan classification based on the
structure of trigger circuits utilizing random forests,’’ in Proc. IEEE 27th Int.
Symp. Line Test. Robust Syst. Design (IOLTS), Jun. 2021, pp. 1–4.
[16] T. Lu, F. Zhou, N. Wu, F. Ge, and B. Zhang, ‘‘Hardware Trojan detec-tion
method for gate-level netlists based on the idea of few-shot learn- ing,’’ in
Proc. IEEE 21st Int. Conf. Commun. Technol. (ICCT), Oct. 2021,pp. 301–305.
[17] S. Li, Y. Zhang, X. Chen, M. Ge, Z. Mao, and J. Yao, ‘‘A XGBoost
based hybrid detection scheme for gate-level hardware Trojan,’’ in Proc. IEEE
9th Joint Int. Inf. Technol. Artif. Intell. Conf. (ITAIC), Dec. 2020,pp. 41–47.
[18] Y. Zhang, S. Li, X. Chen, J. Yao, Z. Mao, J. Yang, and Y. Hua, ‘‘Hybrid multi-
level hardware Trojan detection platform for gate-level netlists based on
XGBoost,’’ IET Comput. Digit. Techn., vol. 16, nos. 2–3, pp. 54–70, Mar.
2022.
[19] P. Zhao and Q. Liu, ‘‘Density-based clustering method for hardware Trojan
detection based on gate-level structural features,’’ in Proc. Asian Hardw.
Oriented Secur. Trust Symp. (Asian HOST), Dec. 2019, pp. 1–4.
[20] R. Sharma, N. K. Valivati, G. K. Sharma, and M. Pattanaik, ‘‘A new hard-ware
Trojan detection technique using class weighted XG-Boost classifier,’’in Proc.
24th Int. Symp. VLSI Design Test (VDAT), Jul. 2020, pp. 1–6.
[21] B. Safaei, A. M. H. Monazzah, M. B. Bafroei, and A. Ejlali, ‘‘Reliability side-
effects in Internet of Things application layer protocols,’’ in Proc. 2ndInt. Conf.
Syst. Rel. Saf. (ICSRS), Dec. 2017, pp. 207–212.