An Intrusion Detection and Defence Mechanism for Securing Autonomous Vehicle CAN Bus Using AI-Based Homomorphic Encryption

Modern technical developments in the automotive industry have focused on modern automobiles [1]. Because of this, the modern complexity of vehicles is rising and providing a greater range of cutting-edge, practical apps that encompass several functions. These characteristics are managed by hundreds of Electronic Control Units (ECUs), which are connected to one another over a Controller Area Network (CAN) bus [2]. Vehicle subsystems are monitored and controlled by the ECU to increase energy efficiency and lessen noise and vibration [3]. In order to safeguard communication from online threats, CAN lacks security features like authentication and encryption. Researchers have shown that in-car networks are seriously vulnerable to security flaws. An attacker can physically access a car through a variety of means, such as manipulating and accessing an Electronic Control Unit (ECU) through faulty designs or sending fake messages into the CAN system [4].

High connectivity and automobile electronics are developing as creative solutions to improve driving convenience [5]. Smart devices and cellular networks are used in vehicle-to-vehicle communication to transmit vital information, especially unsafe road conditions. Sensors in autonomous cars enable vehicle-to-infrastructure communication. Still, the improvements in automobile connectivity are susceptible to outside attacks, such as the absence of authenticating mechanisms in the current CAN messaging frame and the increasing complexity of in-car controller designs [6]. This might cause unwanted movements or failures, threatening the safety of passengers and vehicle cybersecurity.

Cybersecurity in essential contexts, such as automobiles, necessitates high-accuracy intrusion detection systems (IDSs) to avoid false alarms and safeguard passengers, pedestrians, and other drivers from malicious assaults [7]. Real-time reaction is critical for vehicle cybersecurity, but due to time and space limits, in-vehicle systems might not react in an instant. A real-time IDS is required to perform with restricted resources. The CAN bus technology contains technological flaws, rendering it susceptible to network-based assaults [8]. CAN systems have been developed by a number of businesses, including Google, Tesla, the University of Michigan, BMW, Audi, Mercedes Benz, and Baidu. Spoofing and flood attacks, which fall under the categories of passive and active attacks [9].

Physical aspects of CAN protocol security, such as access restrictions and encryption, have received the most attention [10]. However, there is a need for more efficient IDS to address the limitations of CAN bus communications. To identify threats, traditional ML-based IDS algorithms such as Naive Bayes (NB), support vector machines (SVM), decision trees (DT), Multi-Layer Perceptron (MLP), and random forests (RF) are employed. Due to the limited processing capability of traditional ECUs, these algorithms are especially effective in vehicle networks. A systematic framework for detecting and classifying cyberattacks is still inadequate. However, employing a single base classifier, high-degree-of-freedom models may fail to match data distributions effectively. The computational cost of these options makes them unsuitable for use in critical systems like AV system. This proposed method presents a novel IDS for autonomous vehicles by integrating Fully Homomorphic Encryption (FHE) with a conservative physics-informed neural network (cPINN). Unlike traditional IDS approaches, which often require access to unencrypted data and may compromise privacy, the proposed system enables secure and accurate intrusion detection directly on encrypted CAN bus data. By leveraging the strengths of FHE and cPINNs, the system ensures data privacy without sacrificing detection accuracy or computational efficiency. This makes it particularly well-suited for resource-constrained environments typical of autonomous vehicle systems, providing enhanced cybersecurity for in-vehicle CAN bus networks.

Main contributions of the proposed technique are listed as follows:

• Homomorphic encryption-based machine learning is utilized to improve data security in Connected and Autonomous Vehicles (CAVs) by recognizing possible attack data.
• Random drop imputation and Batch Normalization (BN) are utilized as pre-processing techniques to impute the missing values and to standardize the given data.
• The PHATE technique is developed to reduce the dataset dimensionality.
• Homomorphic encryption, named as Cheon, Kim, Kim, and Song (CKKS) based cPINN is utilized to secure the information by identifying the possible attacks in CAVs.
• Modified TrustNet is utilized to mitigate attacks on CAN bus critical communications, ensuring minimal maintenance and protection from potential threats.
• For analyzing the efficiency of the proposed FHE-cPINN model, the performance of accuracy, precision, F1 score are significantly higher than existing approaches.

This paper is structured as follows: Section 2 provides a comprehensive summary of relevant work, including its shortcomings. Section 3 provides a brief overview of the proposed technique and its associated architectures. Section 4 contains the results and a discussion. Section 5 presents the conclusion and future scope.

Modern vehicles are built with ECUs and sophisticated computing systems that enhance communication and comfort. However, greater connection raises security problems since the CAN is vulnerable to hackers. IDSs are crucial for protecting in-vehicle networks. Effective IDSs have been developed using machine learning and deep learning techniques. However, existing Machine Learning (ML)-based IDSs face issues such as poor detection accuracy, delayed real-time response, and inadequate processing power due to network traffic quirks. Therefore, encryption-based supervised ML classification algorithms are essential in solving the challenges with data security in CAVs by determining anticipated messages of attack. The proposed approach for securing and detecting threats against a CAV network is depicted in Figure 1, and its working flow is given in Figure 2. Initially, CAN Bus dataset is collected from real CAN traffic data. Random drop imputation, and BN is utilized as pre-processing techniques to impute the missing values and standardize the input data. Then, dimensionality of the dataset gets reduced by PHATE. These reduced data are trained using homomorphic encryption named as CKKS based on cPINN that secures and diagnose attacks on CAN buses in vehicles. A proposed prediction model can detect whether there is attack or non-attack. Once an attack has been detected, the modified TrustNet is employed to mitigate the attack. Modified TrustNet attempts to protect CAN-bus critical communications from attacks with minimal maintenance. The diagram presents a comprehensive pipeline for detecting attacks in a CAN bus dataset. The process begins with preprocessing, where missing values are addressed using Random Drop Imputation and the data is standardized through BN to ensure consistent scaling across features. Following preprocessing, the high-dimensional dataset undergoes dimensionality reduction using PHATE (Potential of Heat-diffusion for Affinity-based Transition Embedding), which effectively reduces the feature vector while preserving essential data structure. The reduced data is then fed into a classification module that combines homomorphic encryption with a cPINN. This enables secure and privacy-preserving classification of the data into either "attack" or "non-attack" categories. Detected attacks are subsequently handled by a Modified TrustNet framework, which likely further evaluates or mitigates the identified threats, ensuring secure and trustworthy decision-making within the system.

Figure 2. Working flow of the proposed method

3.1 Pre-processing

A crucial step in the data mining process is data pre-processing. Before data is analyzed, it must first be combined, cleaned, and converted. Increasing the data's quality and applicability for the particular data mining activity is the goal of data preparation. In this proposed approach, Random drop imputation (RDI) is used to replace missing values in a raw input dataset, and the data is adjusted within a certain range using BN.

3.1.1 Random drop imputation

An imputation model is explicitly trained by RDI using random drop data, which is produced by arbitrarily eliminating observed values from the time-series data [17]. Let's construct $\tilde{X}=\left\{\tilde{x}_1, \ldots, \tilde{x}_T\right\} \in \mathbb{R}^{T \times D}$ to represent the random drop data, where each element is a vector $\tilde{x}_t=\left\{\tilde{x}_t^1, \ldots, \tilde{x}_t^D\right\} \in \mathbb{R}^D$. Additionally, designate $\widetilde{M}$, a mask whose element $\widetilde{m}_t^d$ is specified as follows, and which indicates the position of both the initially missing values and the randomly dropped data:

$\widetilde{m}_t^d= \begin{cases}0, & {if\ } \tilde{x}_t^d {as\ missing\ value} \\ 1 & {otherwise}\end{cases}$            (1)

The imputation loss, also known as the loss function of RDI, has the following expression:

$\begin{aligned} & L_{\text {impute }}(X, \widetilde{M})=\| X \odot(M-\widetilde{M})-F(\widetilde{X ; \theta}) \odot   (M-\widetilde{M})\left\|_2+\right\| \tilde{X} \odot \widetilde{M}-F(\widetilde{X ; \theta}) \odot \widetilde{M} \|_2\end{aligned}$             (2)

where, $\theta$ is the parameter of an imputation model F. By producing missing values with ground truth, the proposed technique, RDI, enables the explicit training of an imputation model. Additionally, it provides data augmentation by creating random drop data from the source data. Through ensemble learning, the supplemented data is used. Although it has bias difficulties, bootstrap with ensemble learning was employed to overcome unstable and over-fitting issues. However, RDI augments the whole dataset to produce distinct sets, making it possible to apply ensemble learning without bias. The procedure is as follows: $N$ imputation models $F_k$ are trained with various data using the loss function (2), and $N$ random drop data, represented as $\tilde{X}_k$, are generated. The drop data is used to create original data for each model $F_k$, which is then provided to each model Fk that has already been trained for imputation. Since each model's output values are unique, the ensemble model's ultimate output is determined by averaging its $N$ output values.

3.1.2 Batch Normalization

BN is a technique for minimizing the internal covariate shift induced by changes in input signal distribution [18]. Instead of utilizing the statistics of the complete dataset to normalize intermediate representations, a mini-batch is used to lower processing costs. The mini-batch's input characteristics are normalized using the determined statistics as,

$\hat{x}_b=\frac{x_b-\mu_B}{\sqrt{\sigma_B^2+\epsilon}}$           (3)

where, $\epsilon$ is a numerical stability-related tiny constant. The input characteristics, which can be expressed as $B N\left(x_I\right)=\gamma \hat{x}_i+\beta$, are transformed using the learnable parameters $\gamma$ and $\beta$ to further enhance the layer's representation capacity. BN uses the running average of the training mean and variance during the inference stage. In contrast to the static BN, this study uses SNNs to enable temporally-varying parameters in BN, allowing exploration of the temporal properties of BN.

3.2 Dimensionality reduction

Dimensionality reduction is a strategy for reducing the number of features in a dataset while maintaining its original qualities. To develop a model with fewer variables, unnecessary or redundant characteristics, as well as noisy data, are removed. This technique includes a variety of feature selection and data compression strategies employed during pre-processing, all of which reduce high-dimensional spaces to low-dimensional ones. In this work, an approach known as PHATE decreases dimensionality and visualizes data while retaining both local and global data structures.

3.2.1 PHATE

A novel feature analysis viewpoint for deep forest classification is provided by the dimensionality reduction technique known as PHATE, which maintains local relationships between data points while learning general spatial features. In line with other techniques, it reduces the eigenvector dimension to 32 [19]. The nonlinear, unsupervised method known as PHATE maintains local and global data links and accurately represents high-dimensional data sets while combining the advantages of PCA and tsne.

The following are the particular actions.

An expression for the eigenvector of a series is $x_n n, \in 1, \ldots, k$, where ${k}=200$. The Gaussian kernel function quantifies the similarity between $x_a$ and $x_b, a, b \in\{1, \ldots, k\}$, based on the Euclidean distance between them, expressed as $k_z\left(x_a, x_b\right)$.

$k_z\left(x_a, x_b\right)=\exp \left(-\frac{\left\|x_a-x_b\right\|^2}{\epsilon}\right)$             (4)

wherein the neighborhood radius that the kernel function captures is determined by the bandwidth measurement, denoted by $\varepsilon$.

Reduce stress by using metric multidimensional scaling (MDS) measurements $\left(\hat{x}_1, \ldots, \hat{x}_{32}\right)$,

$\begin{aligned} & \operatorname{stress}\left(\hat{x}_1, \ldots, \hat{x}_{32}\right)= \sqrt{\sum_{i, j}\left(\Re_{\left(x_i, x_j\right)}^t-\left\|\hat{x}_i-\hat{x}_j\right\|_J\right)^2 / \sum_{i, j}\left(\Re_{\left(x_i, x_j\right)}^t\right)^2}\end{aligned}$            (5)   

A 32-length fixed-length vector has been generated by capturing the data in the MDS embedding. In order to identify message attacks in CAN and secure the vehicle's network from cyber threats, these reduced dimensionality data are trained using homomorphic encryption, named as CKKS based cPINN.

3.3 Classification

Reduced data serves as input for the classification process, which secures and predicts if the network is attacked or not. A classifier is an algorithm that sorts data into labeled groups or categories of information. The proposed methodology uses homomorphic encryption named as CKKS based cPINN, which has undergone extensive training and evaluation on test and training data.

3.3.1 Homomorphic encryption

For real numbers, a previously explained homomorphic computing algorithm is homomorphic encryption for Arithmetic of Approximate Numbers (HEAAN), sometimes referred to as CKKS [20]. Ring-Learning with Errors (Ring-LWE) employs noise, sometimes known as error e, to rely on FHE methods. This noise is encrypted as part of the payload for security purposes. The payload and noise are combined to generate the plaintext (µ + e), which is then encrypted.

In CKKS, the space between the plaintext and the ciphertext is nearly equal. These components are based on the polynomial rin $R_S=\mathbb{Z}_S(y) / f(y)$, where s is the integer coefficient modulus and $f(y)$ is exponent of polynomials. The polynomials that make up $R_S$ have integer coefficients that are constrained by s . In addition, the degree of $f(y)$ limits their degrees. With $k$ being a power two known as the ring dimension, the formula for $f(y)$ that is most frequently seen in research is $f(y)=y^k+1$. The number ring components changes with each appearance of the CKKS ciphertext and plaintext. The CKKS plaintext has a single ring elemen (polynomial), while the CKKS ciphertext contains two or more ring elements.

CKKS provides encoding and decoding (codec) algorithms that allow a vector of real numbers, or more precisely, complex numbers, to be translated between plaintext space and another vector. Integral operands can be used in combination with integer operations to carry out fixed-point arithmetic operations, which are represented by CKKS.

1. Plain Text Encoding and Decoding

The CKKS investigation is further enhanced by a novel codec technique that converts a vector of complex numbers into an equivalent plaintext item and vice versa. One single plaintext element, $b \in R$, is produced during the encoding of $a \frac{k}{2}$ vector containing complex numbers, $y \in c^{\frac{k}{2}}$.

Decoding was the process of reversing a component of plaintext and yielding a series with complicated numbers. Procedure makes use of Eq. (12). A version of the Fourier transform is the complex canonical embedding, or map π.

$\operatorname{encode}(y, \Delta)=\left\lfloor\Delta . \pi^{-1}(y)\right\rceil$           (6)

decode $(b, \Delta)=\pi\left(\frac{1}{\Delta} \cdot b\right)$         (7)

The relationship between the fixed-point format and the CKKS codec method ought to be clear. In order to decrease the smallest fractional components during encoding, rounding is performed. Scaling is accomplished by multiplying the supplied payload by $\Delta$. In decoding, it's a reverse method. When encoding, multiply by $\Delta$, and when decoding, divide by $\Delta$.

2. Parameters

Three variables were analyzed: $\mathrm{k}, \mathrm{s}$, and $\Delta . R_2$ is used for evenly sampling polynomials $\{-1,0,1\}$ with coefficients of integers in CKKS, whereas $\mathbb{X}$ is used for a distribution that is gaussian discontinuity across R utilizing variables $\mu$ and $\sigma$, constrained by an integer $\beta$.

It has been determined that the present homomorphic encryption norm $[\mathrm{ACC}+18]$ as $(\mu, \sigma, \beta)$ is $\left(0, \frac{8}{\sqrt{2 \pi}} \approx 3.2,[6 \cdot \sigma]=19\right)$.

$R_s$ represents an attire randomized over. When q is fixed, the study finds that an RLWE hardness estimate with $n$ offers adequate security. A ring's coefficient modulus is expressed as $s^{\prime}=\frac{s}{\Delta}$ and its size is decreased as a result of the homomorphic multiplication process, which reduces the ciphertext by $\Delta$.

A connection among ciphertext coefficients, where $1 \leq v \leq V$, and $V$ denotes a recently encoded cipher text level, is the coefficient modulus at level v , shown by $s_v$.

$s_V>s_{V-1}>\cdots>s_1$        (8)

3. Key Generation

An interval-based quadratic with degree multiple variables $\{-1,0,+1\}$ is used to construct the secret key $S_k$. Communication encryption is done using AES 256 bit encryption with randomness and predefined settings. Each of the 14 rounds that comprise the AES-256 encryption key generates a 128bit round key. $\left(P_{k 1}, P_{k 2}\right)$ is a set of polynomials that represents the public key $P K$ that was derived using the given parameters.

$P_{k 1}=\left[-b . S_k+e\right]_{s V}$              (9)

$P_{k 2}=b \leftarrow^g \quad R_{s V}$              (10)

Polynomial computation should be performed modulo $s V$, or more specifically modulo the ring polynomial modulus $y^k+1$, since $P_{k 2}$ is in $R_{s V}$ and e is an unpredictability polynomial that $\mathbb{X}$ chose. This is evident from the symbol $[.]_{s V}$.

4. Encryption and Decryption

Either R (which encrypts the input payload vector) or plaintext message m can be encrypted. To get the ciphertext $c=\left(c_1, c_2\right)$, three shorter randomized polynomials ($g$ from $R_2$ and $e_1$ and $e_2$ from $\mathbb{X}$) should be constructed in $R_{s V}^2$ in the sense described below:

The secret key is evaluated to estimate the length of the message in plaintext during decryption of the provided ciphertext at stage $v$.

$\widehat{m}=\left[c_1+c_2 \cdot S_k\right]_{s V}$           (11)

5. Evaluation of Homomorphic

Addition and multiplication using homomorphism are two kinds of homomorphic operations that are carried out.

1. Eval-Add

Eq. (20) illustrates that input cipher messages just need to be treated to the necessary polynomials. This should be taken into account because the input cipher texts are probably equivalent in terms of level and scale. If not, the ciphertext levels at higher levels must be decreased to match those at lower levels.

$\begin{aligned} & \text { eval-add }\left(c^{(1)}, c^{(2)}\right)=\left(\left[c_1^1+c_1^2\right]_{S V},\left[c_2^1+\right.\right. \left.\left.c_2^2\right]_{S V}\right)=\left(c_1^3, c_2^3\right)=c^3\end{aligned}$          (12)

6. Eval-Mult

Eval-Mult may be calculated using two cipher messages and the following technique:

$\begin{aligned} & \operatorname{eval}-\operatorname{mult}\left(c^{(1)}, c^{(2)}\right)= \left(\left[c_1^1 \cdot c_1^2\right]_{s V},\left[c_1^1 \cdot c_2^2+c_2^1 \cdot c_1^2\right]_{s V},\left[c_2^1 \cdot c_2^2\right]_{s V}\right)= \left(c_1^3, c_2^3, c_3^3\right)=c^3\end{aligned}$           (13)

Relinearization uses $R_{s V}$ for calculation to minimize the size of $c^3$ polynomials in a ciphertext. The product is encrypted using the nonexpanded ciphertext with a squared scale factor of $\Delta^2$. To repeat the operation, two ciphertexts in CKKS can be multiplied with various $\Delta_1$. $\Delta_2$ scale factors. It is necessary to make sure the product does not wrap around or fit inside the datatype in order for this fixed-point arithmetic approach to be practical. It is suggested that for easy representation, a fixed scale factor be used throughout the calculation.

3.3.2 Conservative physics-informed neural network

An expansion of PINN designed to solve the conservation lw is called cPINN. It addresses the problems of PINN's accuracy and expensive training costs. Domains are subdivided into non-overlapping subdomains that take into account various neural network topologies. This makes it possible to choose networks for each subdomain more effectively. After that, the solution in each subdomain is reassembled with the appropriate interface conditions. Because of its unique purpose, which permits different system and scalar architectures in each subdomain, cPINN is still utilized for controlled equations even though it hasn't been studied extensively on hyperbolic conservation laws [21].

Let $N^L: \mathbb{R}^{D_i} \rightarrow \mathbb{R}^{D_a}$ be a neural network based on feed forwarding with $L$ layers and $N_k$ neurons in the kth layer $\left(N_0=D_i\right.$, and $N_L=D_a$. The kth layer $(1<k \leq L)$ weight matrix and bias vector are represented, respectively, by $W^k \in \mathbb{R}^{N_k \times N_{k-1}}$ and $b^k \in \mathbb{R}^{N_k}$. Denoted by $z \in R^{D_i}$ for the input vector and $N^k(z)$ and $N^0(z)=z$ for the output vector at the $\mathrm{k}_{\ldots n}^{\text {th}}$ layer, respectively. The scaling factor $(n)$ and the scalable parameters $n a^k$ are applied layer-wise, with $\Phi$ representing the activation function. Additional parameters $a^k$ change activation function slope in hidden-layers, increasing training speed and contributing to loss function through slope recovery term.

$S\left(a^k\right)=\frac{1}{\frac{1}{L} \sum_{k=1}^L N\left(a^k\right)}$               (14)

where, the operator for the exponential is $N$. The network's ability to learn is improved by such locally adaptable activation functions, particularly in the early training phase. A mathematical instance of this may be provided by contrasting the slope behavior of the static activating technique with the adjustable function of the activation approach. The gradient-dependent dynamics of adaptable activation differ from those of fixed activating by dividing an induced vector by a slope and introducing the predicted seconddegree component. The present study, initialized $n a^k=1, \forall k$, and employed scaling factor $n=5$ for every hidden layer. Figure 3 displays the cPINN algorithm design.

Figure 3. Adaptive activation function-based conservative physics-informed neural network (cPINN) schematic

The groundbreaking technology known as cPINN separates the domain into small sub-domains and enables the use of many neural networks with different topologies to solve PDEs. This method greatly accelerates convergence rates and increases computing efficiency. By pulling them from zero-mean, finite-variance transportation, biases and weights are initialized using the Xavier initialization technique. For each given layer with N nodes, the optimal variance value is 1/N. Methodology also allows for the selection of network hyper-parameters depending on solution regularity, such as optimization method, activation function, depth, or width. This enables the deployment of shallow networks for smooth zones and deep networks for complicated areas.

The purpose of neural network training is to make use of available training data, which can be gained by starting and boundary conditions, properly executed experiments, or high-resolution numerical experiments. The data set is classified as low or high-fidelity depending on measurement inaccuracy and can be utilized to solve PDEs.

Sub-domain optimization technique and loss function:

The cPINN algorithm learns a substitute, $u= u_{⊛̃}$, for predicting the solution u for any input $x_u^i$ given a PDE model and a set of training data, $\left\{u\left(x_u^i\right)\right\}_{i=1}^M, x^i:=\left(x_u^i, y_u^i, t_u^i\right) \sim p(x)$. Typically, one must approximate the density $p(x)$ using the input training data as it is not known a priori in most circumstances. The loss function, which is equipped with interface conditions and resembles the PINN loss in structure, may be specified subdomain-wise. Each sub-domain's loss function for the forward problem is provided by

$\begin{aligned} & \mathcal{L}(\widetilde{\circledast})_{ {pth sub-domain }}=W_{u_p} \times M S E_{u p}+W_{F_p} \times \ M S E_{F_p}+W_1 \times \underbrace{\left(M S E_1+M S E_{u_{a v g}}\right)}_{ {Interface conditions }}\end{aligned}$             (15)

where, $W_{F_p}, W_1$, and $W_{u_p}$ denote the interface, residual, and boundary/initial weights, respectively. Using these classification approaches, attacks and non-attacks in CAN are detected, thereby securing the vehicle network from cyber threats. If an attack has been identified, it is defended by using the modified TrustNet.

3.4 Modified TrustNet

In Cyber-Physical Systems (CPS) networks based on CAN, Modified TrustNet is a secure technique that guards against cyberattacks. High safety standards are achieved without adding any additional information to the communication frames; instead, a conventional CAN bus with security features implemented in software or hardware is used. One way to maintain compliance with CPS's real-time requirements is to fully integrate Modified TrustNet in software [22]. Without requiring any hardware or architectural modifications, it provides minimal overhead protection against masquerade and replay attacks. Supporting all CAN frame IDs specified in the original CAN standard, modified TrustNet offers complete reversibility. Although CAN protocols are used in many applications, they were originally designed as an in-vehicle network. Malicious assaults on open systems that could lead to future issues or compromise the privacy of sensitive data must be addressed. Conventional security methods might not work due to CAN's physical and restricted nature and the requirement for real-time processing by inexpensive, low-performance microcontrollers. Below Figure 4 shows the legacy CAN peripheral is a component of middleware that communicates with TrustNet.

Figure 4. An Electronic Control Unit (ECU) that is based on an antiquated Controller Area Network (CAN) peripheral and uses technology to enable TrustNet

To enable modified TrustNet, which virtualizes the CAN bus and allows for both safe and insecure communication over the same physical channel, both filters and masking on outdated CAN devices need to be adjusted to 0. To maintain safe communication, modified TrustNet requires that reliable ECUs share private keys. Modified TrustNet segment nodes sync on a regular basis, according to the time intervals established by CAN frame transmissions on the CAN bus. Nodes keep broadcasting frames unless a synchronization frame indicates that a fresh period is in effect. The alignment frame is encoded as an unsecured CAN frame, as specified.

• ID:0 × 0
• DLC==0 × 1: Index of the Keys Table determined by the highest entry, duration set to 100 frames

LS portion's keys database index and the MS component's time interval make up the split payload for DLC > 0 × 1.

Arbitrators on the CAN bus favor cycles with dominating bits over negative ones, making it susceptible to attack. By ensuring the encryption process functions properly, using a CAN ID of 0 guards against malicious attacks. CAN controllers provide a separate $T x$ and $R x$ route, allowing applications to send and store CAN frames. The program manages the transmission process using "send-and-forget" functionality. Because Trust-Net encryption operates atomically in a single CAN structure, updating counters and HMAC data, synchronization issues might arise.

Separating the $T x$ and $R x$ procedures and using two separate counters for each direction is the next way to address the characteristics of CAN controllers. To do this, one counter must be added for each tuple of communication nodes and direction, such as $(T x A, R x B)$ and $(R x A, T x B)$. A node must keep as many counters per connection/direction if it is set up to communicate with many (or ALL) other nodes on the CAN bus. After the payload and CAN ID are extracted at reception, the CAN ID is subjected to an XOR operation using ALL valid counters at the RX route. By comparing the latest calculated CAN IDs with the list of valid IDs, XOR computations and search cost are minimized. This method effectively prevents replay attacks. Algorithm 1 provides the pseudo code for the proposed algorithm for securing and detecting attacks against a CAV network.