In-vehicle network intrusion detection systems: a systematic survey of deep learning-based approaches

Research scope and questions

This survey is dedicated to identifying, classifying, and comparing current DL-based IDS schemes for securing In-vehicle networks. Based on this purpose and the analysis of uncovered content in other reviews, the following Research Questions (RQ) were defined.

RQ1 Which DL techniques(methods) are adopted by each IDS scheme?

RQ2 How do different schemes employ DL techniques in different stages of the intrusion detection procedure to achieve better performance?

RQ3 Which datasets are used for IDS training, validating, and testing purposes?

RQ4 What evaluation criteria are applied to measure the performance of the system?

RQ5 What are the merits and demerits of each IDS scheme for recognizing anomalous behavior in considered domains?

Inclusion and exclusion criteria

The articles retrieved from databases were further screened to decide whether they should be included for data extraction and analysis. We follow a two-stage strategy: articles were first submitted to a coarsely screened through the article abstract, followed by a full-text review for detailed quality assessment. The inclusion and exclusion criteria used in this survey are listed in Table 2.

Injection attacks

The injection attack sends packets with malicious data to the network. While the attacker injects malicious messages, the original sender still sends normal messages. From the perspective of the receiver node, the number of received messages is increased. Most of the attacks studied in IVN IDS schemes fall into this category, which can be subdivided into the following:

DoS attack: A DoS attack aims to exhaust the target ECU resources or deplete network bandwidth by sending massive legitimate requests. In the CAN bus, for instance, the adversary utilizes the priority arbitration mechanism to send a large number of high-priority messages (CAN ID:0x000), preventing other ECUs from transmitting their messages (Studnia et al., 2013; Palanca et al., 2017). The probability and catastrophic of a DoS attack are considered very high owing to adversaries with limited prior knowledge can implement this.

Replay attack: The purpose of a replay attack is to record previously propagated messages in the network by eavesdropping and re-transmit them later. For instance, The CAN bus is vulnerable to replay attacks due to a lack of freshness mechanisms. An attacker could access the engine speed signal and performs the replay attack, causing other ECU uses outdated information and impacting vehicle functionalities (Hoppe & Dittman, 2007).

Spoofing attack: The attackers inject fake messages into the network with the purpose of overlaying signals sent by the original ECU and deceiving the receiver (Larson, Nilsson & Jonsson, 2008). For example, the attackers only modify the corresponding bits of the CAN message representing the RPM (revolution per minute) gauge (Martinelli et al., 2017). While modifying a signal with a specific function requires the attacker to have prior knowledge of the communication matrix or execute reverse engineering to the traffics. Usually, this type of modification attack injects a small number of messages, making it challenging to identify.

Fuzzing attack: A fuzzing attack is used to investigate the impacts of different packets on the ECUs by transmitting messages with random protocol headers and payload into the network at a high frequency (Martinelli et al., 2017; Fowler et al., 2019). Unlike DoS attacks, inject packets may conform to normal traffic characteristics, causing the receiver node to use the information in the malicious packets with unexpected results.

Suspension attack

A suspension attack aims to stop some or all message transmission of the target ECUs, thus affecting the functionality of other ECUs that rely on this continuously updated information (Verma et al., 2020a). For example, the adversary can disturb the CAN frame with a particular ID to generate an error frame (e.g., stuff error), which violates the CAN protocol specification, causing the target node to enter an error state. Besides, executing this attack on SOME/IP can intercept and discard the information via a man-in-the-middle (Zelle et al., 2021).

Masquerade attack

The attacker first suspends a specific message and then transmits the spoofed packets with the same transmit interval, message format, and payload value range (Iehira, Inoue & Ishida, 2018). Usually, masquerade attacks do not change the amount of traffic in a period, which is more challenging to identify than injection attacks. The IDS scheme needs to check the reasonableness of the packet payload to detect these sophisticated attacks.

Here, we use the concepts of Timing Transparent (T.T.) and Timing Opaque (T.O.) provided by Verma et al. (2020a), aiming to map these attack categories to different deep learning-based intrusion detection techniques. A Timing Transparent attack means it changes the frequency of network traffic, specifically the packet interval decrease/increase. In comparison, the Timing Opaque attack is defined as it does not disrupt network traffic timing or distributions. For example, the Timing Transparent attack, such as DoS or fuzzing attack, is detectable by the scheme using the CAN ID sequences. On the other hand, the Timing Opaque attack needs a more sophisticated scheme by checking payload data. In Table 3, we itemize the features, results, and types (T.T. or T.O.) of the discussed IVN attacks for comparison and reference.

Autoencoder

AE is the common DL technique applied in IDS research for dimensionality reduction and anomaly detection in an unsupervised way (Hinton & Salakhutdinov, 2006). It generally consists of input and output layers with the same dimensions, an encoding neural network, a decoding neural network, and a latent space. AE works on matching the decoder output as close to the encoder input as possible by performing representation learning, a process called reconstruction. Therefore, It is trained to minimize the reconstruction error between the input and the reconstruction sequence to obtain a low-dimensional abstraction representing high-dimensional data. Then, an input traffic sequence is considered abnormal if its reconstruction error is larger than the specified threshold. This subsection discusses the AE-based IDS schemes (Alqahtani & Kumar, 2022; Cheng, Han & Liu, 2023; Alkhatib et al., 2022) implemented for the in-vehicle network.

Lin et al. (2020) proposed an efficient DL-based IDS scheme using the deep denoising autoencoder, which can learn latent sequential patterns of CAN data. This scheme adds an evolutionary-based optimization algorithm to tune the proposed model parameters, aiming to maximize the performance and efficiency of the model against malicious attacks. This solution is evaluated on the three datasets, including two popular anomaly detection datasets in the area of the CAN bus. They achieved high precision and F1-score, demonstrating that this model is robust and reliable in detecting DoS attacks, fuzzing attacks, and impersonation attacks.

Wei et al. (2023) introduced a NIDS approach, denoted as AMAEID, for CAN bus real-time intrusion detection using the denoising autoencoder and attention mechanism. The hexadecimal CAN payload was converted into binary format in the data preprocessing stage. Then it superimposed the noise conforming to the Gaussian distribution, which benefited the model in terms of interference resistance and generalization. Finally, a single-layer fully-connected network is used to obtain the final prediction value, indicating whether the message is normal or abnormal. The authors validated their model using OTIDS dataset and demonstrated that it achieves high performance. However, this scheme is only evaluated with traditional machine learning methods (e.g., DT, KNN, SVM), and it is not analyzed with other literature using AE to justify the achieved results.

Lokman et al. (2019) proposed an IDS approach by applying deep contractive autoencoders (DCAEs) to extract flow-based features from network traffic data. This scheme imposes different penalty terms on the loss function representation, ensuring the network is robust to minor changes, noise, and missing inputs in the training samples. This scheme is evaluated using real CAN data collected from three brand vehicles through OBD-II. The results are compared with stacked sparsed autoencoders and denoising autoencoders.

The NIDS scheme proposed in Ashraf et al. (2021) utilized the long short-term memory (LSTM) autoencoder algorithm on dealing with attack events that occur at the central network gateways of vehicles. This scheme executes binary classification on incoming traffic. Statistical features of the mean and standard deviation computed for messages within a specific time window are transformed by the normalized likelihood sequence and fed into the network. The author validated their method using the UNSW-NB15 dataset for external network communication and the car hacking dataset for in-vehicle communication. The experimental results on two datasets demonstrate that the scheme achieves 99% and 98% accuracy, respectively.

Longari et al. (2021) also proposed an IDS method by applying LSTM autoencoders to extract latent features from malicious behaviors. The proposed model creates a time series of CAN payload for each CAN ID and learns to update network hyperparameters to minimize the Mahalanobis distance between the reconstructed sequence and the original sequence. Larger reconstruction errors will be flagged as potential exceptions. The model is evaluated using collected CAN frames during vehicles driving in different conditions, such as cities and highways. The authors indicate that the proposed method can effectively perform binary classification of intrusion data by performing the required experiments. However, CANnolo has conspicuous detection time consumption and is unsuitable for strong real-time systems.

Kukkala, Thiruloga & Pasricha (2020) proposed a gated recurrent unit (GRU) autoencoder network for detecting intrusive and anomalous flow-based data. Compared to LSTM-based autoencoders, the training process of this scheme consumes less time. The author designed this IDS approach focusing on monitoring the CAN message signal, so the method is protocol agnostic and not limited to the CAN bus protocol. In addition, the authors propose an intrusion score metric to measure the degree of deviation from normal system behavior to determine whether a signal is normal or abnormal. They evaluated their model using the synCAN dataset published by Bosch and analyzed the memory footprint, inference time, and scalability.

Hoang & Kim (2022) introduced a semi-supervised learning-based IDS approach by incorporating convolutional autoencoder and GAN. This scheme trains a classifier in three separate stages. In its first stage, it uses unlabeled data to train the autoencoder part. In its next phase, the encoder and discriminator are trained together. Then update the encoder again with the labeled data by minimizing the cross-entropy loss. This IDS scheme is tested using public datasets. Experimental results demonstrate that feeding only 40% of labeled data to the model achieves the lowest error rate of 0.1% compared to other supervised methods. The properties of AE-based IDS solutions described in this subsection are compared in Table 5.

RNN

Recurrent neural networks (RNNs) extend the capacities of the traditional feed-forward neural network with its ability to learn sequential data over timesteps, especially suitable for sequences that are not independent of each other or involve contextual associations. It consists of the input, hidden, and output units, where the output of the hidden unit relies on its input at the current time step and the output of the earlier time step. The network parameters are updated through the back-propagation over time (BPTT) algorithm. RNNs may provide a good idea for intrusion detection by predicting the upcoming signals of the near future received frames. The detector then regarded the unexpected deviation from the actual signal as an intrusion. However, RNNs usually only can handle finite length sequences and will encounter the vanishing gradient problem if the output at any given timestep depends on inputs much earlier. Different RNN variants like LSTM and GRU are presented to address these issues.

Several RNN-based in-vehicle intrusion detection approaches, such as (Ma et al., 2022; Tanksale, 2020; Tanksale, 2021; Desta et al., 2020a; Desta et al., 2020b; Tariq et al., 2019; Mansourian et al., 2023; Al-Jarrah et al., 2023), are proposed in the literature, which will be discussed in this section. RNN-based IDS was first applied by Taylor, Leblanc & Japkowicz (2016) to detect malicious behavior on the CAN bus. The classifier works by learning to predict the next frame according to the message previously sent on the bus. The authors validated their anomaly detection method using the abnormal data synthesized by adding CAN frames, deleting expected frames, and modifying data contents. Experimental results show that this scheme can successfully flag suspicious anomalies. Nevertheless, a separate detector was trained for each CAN ID without considering the dependencies between different IDs.

Qin, Yan & Ji (2021) presented a NIDS established on the LSTM technique, identifying abnormals and keeping a long-term memory of them. This research focuses on time series prediction, calculating the loss between the predicted content and that transmitted on the bus at the next time step. Then, it is flagged whether the packet is abnormal or normal based on the deviation of the loss value from the threshold. They evaluated the performance of the proposed model on multi-format data input and five loss functions. Nevertheless, this method is not tested using other publicly available datasets.

Zhu et al. (2019) proposed a mobile edge-assisted anomaly detection method for IVNs based on LSTM. It alleviates the problem of insufficient computing resources for onboard devices when deploying deep learning-based anomaly detection systems to vehicles. The model consists of a local common hidden layer computation unit and two single LSTM neural networks that separately process time-based and data-based features. The necessary experiments indicated that with the assistance of the mobile edge, their IDS approach reaches 0.9 detection accuracy, and a CAN message detection is completed in an average of 0.61 ms. However, applying the MEC server introduces indeterminate network latency when delivering CAN bus messages to the MEC server.

Alkhatib, Ghauch & Danger (2021) present an RNN-based IDS to handle the offline intrusion detection problem on SOME/IP application layer protocol. They deal with class imbalance using the adaptive weighting technique that weights samples in rare classes at high costs. The experiments are conducted on the datasets constructed using a packets generator implemented in python language. Results validate that the proposed RNN-IDS provides good results with F1 Scores and AUC values of each intrusion type. However, this scheme merely evaluates attacks that deviate from the protocol specification in the inter-device communication session. In another work (Luo et al., 2023), the authors conducted a multi-layered IDS with GRU-based modules that could simultaneously detect anomalies on SOME/IP header and payload.An LSTM-based IDS was presented by Hossain et al. (2020) in the binary and multi-class classification of the CAN bus DoS attacks, fuzzing attacks, and spoofing attacks. Several layer LSTM networks are utilized to identify cybersecurity attacks, and their outputs are fed into a softmax function. They also investigated the impact of different learning rates, activation functions, loss functions, the number of neurons, and optimizers on model performance with Vanilla LSTM and Stacked LSTM, respectively. This model is evaluated on the self-collected dataset and Survival Analysis datasets, and the results indicate the effectiveness of their approach by comparing it with the Survival Analysis method.

Tariq, Lee & Woo (2020) proposed CAN-ADF, a hybrid DL-based IDS model for detecting anomalies and intrusions formulating as a multiclass classification problem. More specifically, this method uses heuristics-based and LSTM-based methods to identify attacks in in-vehicle network traffic. This combination is because diverse anomalies and attacks arise with varying levels of detection difficulty. Methods with different algorithms may be good at detecting different types of attacks. They evaluate the detection performance of the proposed model by collecting 7,875,791 CAN messages from KIA Soul and Hyundai Sonata. The experimental results demonstrate that the proposed scheme has an average accuracy of 99.45% on multi-classification problems.

Khan et al. (2021) introduced a multi-stage intrusion detection method by combining a normal state-based bloom filter and bidirectional long short term memory (BiLSTM) to discover attacks from communication networks of AVs efficiently. They indicated that bloom-filter with the advantages of efficient memory usage, light-weighted, and constant lookup time to recognize whether network data matches the normal communication database before being fed into the neural network. The authors evaluated their intrusion detection model. The results show that the model achieves an accuracy of 99.11% on the car-hacking dataset and 98.88% on the UNSW NB-15 dataset, representing CAN bus traffic and IP-based network traffic, respectively.

Driss et al (2022) proposed a cyberattack detection framework for vehicular sensor networks that applies GRU. This IDS scheme considers the relationship between previous and present data, which can improve detection capability and rate. The authors employ a group of GRU with a Random Forest (RF)-based ensemble unit to solve the problem of privacy data and resource limitation of intelligent sensor networks. They tested their anomaly detection approach and applied metrics such as accuracy, recall, precision, and F1 score. Experimental results also indicated that their approach outperformed other classifiers. Table 6 presents the various aspects of the RNN-based IDS approaches.

CNN

Convolutional neural networks (CNNs) are specialized neural networks destined to perform image processing and analysis. It generally consists of a stack of input, convolutional, pooling, and fully connected layers. Compared with other deep learning networks under the same network depth, CNN requires fewer parameters, reduces complexity, and speeds up the learning process since CNN adopts local connections and weight sharing instead of traditional fully connected networks. Indeed, It can be usually applied in traffic intrusion detection to perform supervised feature extraction and classification tasks by organizing network traffic into multiple arrays, such as the 1D array of CAN IDs and signals.

This section investigates the CNN-based IDS schemes presented in the literature, such as Lin et al. (2022), Ahmed, Jeon & Ahmad (2023) and Taslimasa et al. (2023). Song, Woo & Kim (2020) presented an IDS scheme, which preprocesses the CAN bus traffic data and then models the temporal sequential patterns using deep convolutional neural networks. This scheme directly converted the bit-stream data of the CAN bus into a grid-like structure, more specifically, building the 29-bit IDs of consecutive 29 frames into a 29*29 matrix as features of the CNN. The IDS scheme adopted the Inception-ResNet, which performs well in natural image classification tasks, to detect abnormal data by reducing the original model components and tuning the network hyperparameters based on the difference in feature dimensions. By conducting the experiments, the authors indicated that this model achieves better detection results compared with LSTM, ANN, and as well as SVM, KNN, NB, and decision tree. In particular, the IDS shows better performance when detecting fuzzy attacks. Moreover, the authors conducted offline testing on devices with GPUs and only CPUs. The results exhibited that the proposed model is challenging to employ online detection in existing automobiles by calculating the detection latency of the processing packets.

Desta et al. (2022) proposed Rec-CNN, a CNN-based IDS technique for investigating network flow data for abnormal activities. This method applies the categorical recurrence plots algorithm, which creates images from time-series data on the CAN bus. Unlike constructing images in Song, Woo & Kim (2020), the authors utilize recurrence plots to capture the temporal dependencies in the sequence of arbitration IDs. They compare the proposed scheme with Song, Woo & Kim (2020), and the experimental results indicate that it achieves better detection performance with comparable execution time.

Jeong et al. (2020) designed a CNN-based IDS to identify the legitimacy of communication nodes, analyzing the channel characteristics between a specific pair of transmitters and receivers. This model takes the bus differential voltage signal as input, outputs the probability that the received message is transmitted from each node, then identifies the node with the highest value as the predicted message source. Then, the detector compares the predicted source with the actual source node number obtained from the mapping table, and the inconsistency is considered abnormal. Simultaneously, the detector continuously transmits dominant bits to generate error frames to interrupt the transmission of abnormal messages. However, if the detector occurs a false alarm, it may seriously block the transmission of normal messages and reduce the CAN bus network performance.

Jeong et al. (2021) introduced a NIDS using CNN to detect AVTP packet replay attacks in automotive Ethernet-based networks. This scheme extracts features from the first 58 bytes of the AVTP stream, inferring the stream AVTPDU is benign or not. The authors captured real AVTP packets as training and testing datasets by building a physical testbed based on BroadR-Reach. Furthermore, they also measure the detection latency on hardware devices such as Google collab, macintosh, Jetson TX2, and Raspberry Pi 3. Results validate that this scheme is efficacious for binary classifications of abnormal traffic.

Han, Kwak & Kim (2023) presented the TOW-IDS solution for heterogeneous in-vehicle networks, which detects the abnormalities of automotive ethernet. This scheme constructed a DCNN model by adjusting the ResNet model as the classification algorithm and furthermore incorporated wavelet transform to reduce the image data size. In consideration of the length difference in different protocol packets (such as CAN, gPTP, and AVB), the authors padded and trimmed the packets to satisfy the computation condition during the data preprocessing process. They evaluated the performance of their approach in identifying frame injection, PTP sync, MAC Flooding, CAN DoS, and CAN replay attacks on self-collected datasets. Experimental results demonstrated that the TOW-IDS consumes less time in detecting network anomalies than default ResNet and EfficientNet methods. In Table 7, the features of discussed CNN-based IDS schemes are listed.

GAN

GAN is a deep generative model developed by Mirza & Osindero (2014), which consists of two ANN networks, namely a generative network and a discriminate network. The generative network learns the distributions of the real data and attempts to generate samples with the same characteristics to confuse the discriminator, which in turn endeavors to distinguish the real data from the generated. Researchers usually utilize GAN models to alleviate the problem of insufficient attack samples in intrusion detection.

Seo, Song & Kim (2018) presented a GAN-based IDS for in-vehicle networks to identify normal and attack traffic for the first time. The generator of the GAN model is composed of CNN, and the discriminator is comprised of DNN. They grouped CAN IDs in the order of sending time and converted them into sample images after one-hot encoding. The authors evaluated their scheme for detecting DoS attacks, fuzzy attacks, and spoofing attacks with the car hacking dataset.

Yang et al. (2021) attempted to train the proposed IDS scheme using the generated CAN images constructed with frames interval time, CAN ID, and DLC. Due to employing a different encoding algorithm compared with (Seo, Song & Kim, 2018), this scheme reduces the input data dimension and the number of network layers and neurons. Furthermore, this scheme applies a sparse enhancement training method benefiting the discriminator to correct its learning direction and accelerate the training convergence, thereby overcoming the poor performance of single GAN. The authors compared the achieved results against (Seo, Song & Kim, 2018) using metrics such as accuracy, recall, precision, and F1 score. The evaluation results outperform related works in terms of precision and accuracy.

Xie et al. (2021) extend the ability of the discriminator to detect tampering attacks by introducing CAN communication matrix that indicates the signal maximum and minimum range. In the data preprocessing stage, the CAN messages are divided into five types according to the transmission mode, effectively reducing discriminator misjudgment. In addition, Considering that the frames sequence is not fixed in the driving environment, the author uses 64 consecutive messages of the same ID to construct the CAN image instead of directly using traffic data that may include different CAN IDs. Experimental results illustrate that their IDS approach could enhance detection performance.

Zhao et al. (2022) proposed an IDS model for implementing multi-classification of CAN bus frames using the auxiliary classifier generative adversarial network (ACGAN). They propose and evaluate the performance of four different variants and indicate that the two-stage classification architecture of ACGAN combined with out-of-distribution works best. ACGAN assigns fine-grained labels to known attacks in the first stage. Out-of-distribution samples are passed into the binary real-fake classifier to perform unknown attack classification. This scheme is tested on public datasets, and the results indicate that the proposed scheme reaches a higher classification accuracy with lower resource overhead. It is suitable for employment in resource-constrained in-vehicle environments. As shown in Table 8, some effective GAN-based IDS models are compared.

DNN

DNNs can be considered an ANN structure with additional depth. It increases the number of hidden layers, enhancing the abstraction capabilities of the model to learn complex nonlinear relationships. DNNs are typical feed-forward networks where data flows unidirectionally from the input layer to the output layer.

This section investigates the articles that utilized DNN techniques to detect in-vehicle network intrusions. Zhang et al. (2019) proposed an in-vehicle IDS based on DNN approach, which attempts to detect spoofing and replay attacks by extracting vehicle behavior features from communication traffics. The author adopts two enhanced BP algorithms, gradient descent with momentum (GDM) and gradient descent with momentum adaptive gain (GDM/AG), to speed up the iterative update process of the network. They collected nearly 300,000 CAN bus messages using the open-source software BusMaster and evaluated this scheme in terms of accuracy, TPR, and FPR. Experiments indicate that the proposed model improves a great deal about the internal cybersecurity of the onboard system.

Cuzzocrea, Mercaldo & Martinelli (2020) employed an MLP-based classifier to recognize DoS attacks, fuzzy attacks, and spoofing attacks on the CAN bus. The proposed method is assessed through experiments on the car-hacking dataset, where data field features are selected. After performing classification experiments using MLPs with multiple hidden layers, they indicated that the MLP network with 1 and 3 hidden layers obtained the best results.

Zhang & Ma (2022) presented a NIDS benefiting from both ruled-based and DNN-based approaches. This ensemble uses deep learning techniques to achieve high detection accuracy while reducing the computational requirements through offloading traffic with a rule-based module. Before conducting detection using the DNN classifier with five hidden layers, the model uses the pre-defined rules to catch malicious messages such as invalid ID, time intervals, and DLC. The authors selected five features with high importance scores in feature engineering for training instead of using the raw data directly. They tested their system on the datasets originating from four vehicles CAN bus data and achieved a binary classification accuracy of 99.8%. However, the model is not evaluated by employing other public datasets to verify its claimed results further. The characteristics of the DNN-based IDS schemes studied in this part are given in Table 9.

DBN

DBNs are probabilistic generative models by stacking several restricted Boltzmann machines (RBM) in layers. In DBN, The network connection of the upper two layers is undirected, and the other layers are directed. DBN is pre-trained using an unsupervised greedy layer-wise learning method followed by learning valuable features using the supervised fine-tuning approach. For IDS, DBNs are usually used to initialize network parameters, improving anomaly detection accuracy. Besides, DBNs are applied to solve ANNs training problems, such as slow training and needing extensive labeled data.

The IDS schemes of this subsection incorporate DBN for in-vehicle communication abnormal classifying. Kang & Kang (2016b) presented an IDS method based on a deep belief neural network. First, they applied DBN to pre-train the neural network parameters intending to initialize the model, followed by the parameters tuned through supervised learning to obtain a better classification performance. The authors utilized all the DATA field for generating the feature. This approach has low computation complexity in the decision, but the training phase is time-consuming. In another article Kang & Kang (2016a), the authors trained and evaluated the proposed DBN-based approach using simulated vehicular network communication called Open Car Testbed and Network Experiments (OCTANE) and simulated the attacker tampering with the tire pressure monitoring signal. They demonstrated that the detection accuracy of this approach is enhanced compared with feed-forward ANN. The characteristics of these two IDS schemes are shown in Table 10.

Hybrid

This section discusses several IDS schemes that combine two or more deep learning algorithms in detecting intrusions procedure.

Other

This section describes some other schemes for detecting intrusion attacks on the IVNs.

Temporal convolutional network (TCN): Shi et al. (2021) presented an anomaly detection system based on the TCN for detecting in-vehicle network attacks. The scheme predicts the sequences by extracting latent features of normal data, flags intrusion when an unknown sequence occurs and further locates abnormal points. Besides, this model can ensure that future information cannot be used to predict the past, which benefits from the TCN network introducing causal convolution that the output at the current time step originates from the convolution of current as well as historical information. The authors conducted evaluations using OTIDS datasets in terms of accuracy, FPR, and TPR. Experimental results indicate that the proposed model performs better in fuzzing and DoS attacks. Also, Thiruloga, Kukkala & Pasricha (2022) presented TENET, an anomaly detection scheme that learns payload-level characteristics using TCN with the integrated attention mechanism.

Complex valued neural network (CVNN): Han, Cheng & Ma (2021) proposed an in-vehicle anomaly detection scheme using the CVNN combined with an encoder and attention mechanisms. Benefiting from the progress of CVNN in model safety, this model can alleviate the privacy disclosure problem caused by massive information collection for training deep learning-based models. The authors utilize a WGAN-based encoder to encrypt intermediate-layer features, followed by a CVNN analysis of in-depth features. The corresponding experimental results demonstrate that this model can protect the IDS while achieving 98% accuracy, and adversaries cannot extrapolate input information from the various features obtained. However, the model efficiency is at average levels compared to other DL-based schemes.

Transfer learning: Mehedi et al. (2021) presented a deep transfer learning-based P-LeNet method for IVN in order to improve the performance of learners on target domains by transferring the knowledge contained in source domains to the target domains. The experimental performance reveals that this model achieved an overall accuracy of 98.1%, which is higher than the other multiple traditional machine learning, deep learning, and deep transfer learning models. In addition, Wei et al. (2022a) proposed a domain adversarial neural network-based IDS consisting of a feature extractor, a label predictor, and a domain classifier. This model learns common feature representations from different but related domains, enabling it can learn the essential features of attacks to cope with the emergence of variant attacks. Furthermore, Tariq et al. (2020) introduced CANTransfer, a convolutional LSTM-based intrusion detection scheme for CAN bus using transfer learning technology. The properties of these IDS models are indicated in Table 12.

Summary

Deep learning has great deployment potential in securing the in-vehicle network. In particular, the detection accuracy of abnormal traffic has reached a high level. For example, several schemes (Kukkala, Thiruloga & Pasricha, 2020; Tariq, Lee & Woo, 2020; Desta et al., 2022) have achieved close to 100% accuracy. However, the classification accuracy and computational complexity of intrusion detection models are in a trade-off relationship. The model with higher inference delay and hardware resource consumption is considered more complex, which negatively impacts the application of an IDS in resource-constrained environments in reality. The urgent challenge to be solved for the DL-based IDSs is reducing their inference delay and hardware resource consumption if the available computing power is poor or an IDS runs on a limited embedded environment of IVNs. In general, each deep learning technique has its own advantages and limitations when applied to the field of intrusion detection issues, as summarised in Table 13. The application of AEs, as an unsupervised learning technique, to intrusion detection schemes is generally used to reduce traffic data dimensions and filter noise. Furthermore, it is also used as a classifier by comparing the deviation between the reconstruction error and the preset threshold, that is, the input sequence is considered abnormal when the error is greater than the threshold. Although it can efficiently compress samples feature but is only available for similar ones and requires a long training time. In contrast, RNNs are an excellent tool since its adequate historical memory for processing temporal sequence data. However, it encounters the problem of gradient vanishing and exploding as the error accumulates over long terms. While different variants like LSTM and GRU are presented to address these issues, it also suffers from overfitting and high memory consumption. CNNs are known for their powerful image classification capabilities. When using CNNs to construct IDS, we need to consider the hardware resources of the on-board equipment since CNN-based models rely more on the computing power provided by the GPUs. Also, it pertains to supervised learning and requires a large amount of labeled data. The main contribution of GANs is that they can alleviate the problem of imbalanced samples by generating synthetic ones, but they also introduce additional operations to check their validity. Moreover, training a GAN model is a challenging task. As for DNNs, they have the high abstraction ability to learn complex nonlinear relationships, but with the increase of hidden layers, a large amount of sample data is required to train the network parameters. Another deep learning technique that can be used to design intrusion detection models is DBNs which help to initialize network parameters. Also, it needs a long initialization phase.