A multi-feature intelligent physical layer authentication method for underwater acoustic networks

Abstract

The application provides a kind of underwater acoustic network multi-feature intelligent physical layer authentication method, network initialization stage is in the case without attack node to build CIR library related channel feature training set, and train the optimal authentication decision boundary for transmission and authentication stage, transmission and authentication stage utilize the optimal decision boundary trained for each legal neighbor in network initialization stage, authenticate newly received message, the purpose is to effectively improve the authentication rate and attack node detection rate under time-varying underwater acoustic channel.The application makes full use of the space-time variation characteristics of underwater acoustic link, provides a new way for high accuracy authentication under the condition of lack of illegal channel prior knowledge;Effectively overcome the influence of dynamic time-varying underwater acoustic channel, thereby further improve the accuracy of physical layer authentication algorithm;Not only can each node in underwater acoustic network realize distributed fraud attack detection, but also can greatly reduce the authentication overhead and network energy consumption generated in underwater acoustic network.

Description

Technical Field

[0001] This invention relates to the field of underwater acoustic network security technology, specifically to a multi-feature intelligent physical layer authentication method for underwater acoustic networks. Background Technology

[0002] In recent years, marine resources and maritime rights have received increasing attention from various countries. Underwater acoustic networks, as a key execution unit, provide crucial support for numerous security applications. However, due to the complex marine environment and the broadcast characteristics of underwater acoustic channels, underwater acoustic networks face significant security threats. Various malicious attacks can compromise the integrity and confidentiality of transmitted data, and even paralyze the entire underwater acoustic network. Authentication technology is the first line of defense for ensuring the security of underwater acoustic networks. Designing effective node authentication mechanisms has become a research hotspot in the field of underwater acoustic network security and is also an important aspect of achieving networked underwater acoustic countermeasures.

[0003] Traditional cryptographic-based upper-layer authentication mechanisms achieve high security at the cost of significant computational and communication overhead, which is intolerable in resource-constrained underwater acoustic networks. Therefore, physical layer authentication technologies, characterized by lightweightness and high reliability, offer a glimmer of hope for addressing network security threats in underwater acoustic networks. ZL202111022616.5 proposes a Time-Reversed Physical Layer Authentication (TRPLA) method. This method utilizes the resonant intensity on the channel impulse response formed by the time-reversal process to spatially focus multipath energy in the underwater acoustic channel. Combining the time-reversed resonant intensity with the Neyman-Pearson criterion effectively detects spoofing attack nodes in the underwater acoustic network. However, the use of a single physical layer feature limits authentication accuracy to some extent. Roee Diamant et al. proposed a joint authentication scheme that fuses multiple features of the underwater acoustic channel collected by a group of trusted nodes to form a strong authentication, and uses a likelihood ratio test to make the final authentication decision. However, this scheme inevitably involves frequent information exchange between nodes, leading to a significant increase in communication costs.

[0004] To further improve physical layer authentication performance, machine learning techniques can efficiently extract features and analyze patterns from physical layer samples. However, unlike terrestrial wireless networks, research on machine learning-based authentication techniques for the more challenging underwater acoustic networks is limited. Specifically, the following two scientific problems urgently need to be addressed in the design of physical layer authentication schemes for underwater acoustic networks: First, underwater acoustic transmission suffers from more significant time- and space-varying characteristics than terrestrial wireless transmission, making the design of authentication schemes that can tolerate dynamic and unreliable underwater acoustic channels extremely challenging; second, underwater acoustic networks have very limited available bandwidth and energy resources. The typical transmit power of an underwater acoustic modem is approximately 30 watts, far exceeding the transmit power of a terrestrial wireless sensor (approximately 80 milliwatts). Considering the enormous transmit power of underwater acoustic modems, frequent message transmission between network nodes for authentication decisions is impractical. Therefore, providing high-performance physical layer security authentication for underwater acoustic network nodes with low authentication overhead and low network energy consumption is particularly important. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention provides a multi-feature intelligent physical layer authentication method for underwater acoustic networks. Addressing the design challenges of authentication algorithms under time-varying, space-varying, and unreliable underwater acoustic channels, this invention proposes a method for constructing a time-resistant multi-feature training set related to the Channel Impulse Response (CIR) library, and utilizes machine learning algorithms to construct an authentication model and deeply analyze features, aiming to improve the accuracy and stability of the physical layer authentication algorithm. To address the challenge of achieving efficient authentication in resource-constrained underwater acoustic network scenarios, this invention utilizes a small-sample underwater acoustic channel multi-feature training set and machine learning algorithms, aiming to achieve high accuracy, low overhead, and low network energy consumption in the physical layer authentication performance of underwater acoustic network nodes. The implementation of this invention does not require obtaining prior information about attackers, providing a new approach to authentication decisions in the absence of prior knowledge of illegal channels.

[0006] The steps of the technical solution adopted by the present invention to solve its technical problem are as follows:

[0007] Step 1: Network initialization phase;

[0008] There is a valid receiving node Alice, and Alice's valid neighbor node is Bob. i 1≤i≤N, where N is the number of valid neighbor nodes Bob. i The number of nodes, all legal nodes are anchored at different depths in the ocean and will drift with ocean currents; nodes Alice and Bob i The underwater acoustic link between them is denoted as AB. i ;

[0009] Step 1.1: Create the CIR library;

[0010] In t mAt that moment, Alice received a message from Bob. i The probe signal, 1≤i≤N, is used to estimate the link AB using the copy correlation method. i At this moment in, For t m Time Link AB i The complex amplitude of the l-th path of the CIR, 0 ≤ l ≤ L-1, where L is the total number of channel paths; to cope with the influence of time-varying underwater acoustic channels, Alice sets the complex amplitude of each of her legal neighbor nodes Bob. i Each establishes a CIR link library M is link AB i The size of the CIR library, Record link AB i CIR sequences at different times were used to capture patterns of CIR variation over time.

[0011] Step 1.2: Construct positive training samples based on CIR library-related features;

[0012] In each legal neighbor node Bob i After establishing the CIR library, in order to construct information about the legal neighbor node Bob i First, Alice uses the link AB to obtain positive training samples. i CIR library any one of them Together with the remaining M-1 CIRs, calculate the multi-features related to the CIR library. F is the number of features, 1≤i≤N, 1≤s≤M;

[0013] Next, Alice utilizes multiple features related to the CIR library. Constitute positive training samples Right now:

[0014]

[0015] in, It is training data (feature vectors). Tags;

[0016] Step 1.3: Construct negative training samples based on CIR library-related features;

[0017] First, Alice uses link AB j CIR library In and link AB i CIR library Given M CIRs, 1≤s≤M, 1≤i≤N, j≠i, calculate the multi-features related to the CIR library. F is the number of features;

[0018] Next, Alice utilizes multiple features related to the CIR library. Constructing negative training samples Right now:

[0019]

[0020] in, It is training data (feature vectors). Tags;

[0021] Step 1.4: Model training;

[0022] First, Alice uses the positive and negative training samples obtained in steps 1.2 and 1.3 to identify her legitimate neighbor node Bob. i Building the training set in, This indicates that Bob is certified. i The q-th training sample is constructed; then, Alice inputs the training set into the SVM model for training, thereby obtaining Bob. i Optimal authentication boundary Right now:

[0023]

[0024]

[0025] in, Bob is certified i The normal vector of the optimal authentication boundary. Bob is certified i The Lagrange multiplier of the q-th training sample, φ(.) denotes the kernel function, b i Bob is certified i The bias term of the optimal authentication boundary. Bob is certified i The v-th unbounded support vector, 0 <v<|V i |,V i Bob is certified i The set of unbounded support vectors;

[0026] Step 2: Transmission and Authentication Phase;

[0027] There is a valid receiving node Alice, and Alice's valid neighbor node Bob. i 1≤i≤N, where N is the number of valid neighbor nodes Bob. i The number of attacking nodes Eve j1≤j≤K, where K is the number of attacking nodes; attacking node Eve j Use with Bob i The same node identity (ID) sends data packets with the aim of gaining network access, thereby launching a series of malicious attacks to destroy the normally functioning underwater acoustic network;

[0028] Step 2.1: For a newly received message at time t', Alice obtains the ID of the sending node claimed by the message. i And estimate the current CIRh i' ;

[0029] Step 2.2: Alice selects the corresponding node based on the sending node's IDi. Based on formulas (7) and (8), using CIRh t' and CIR library Calculate multiple features related to the CIR library And form the feature vector

[0030] Step 2.3: Alice uses the optimal authentication boundary obtained in Step 1.4 Calculate eigenvectors Authentication decision value Where sgn(.) is the sign function and φ(.) is the kernel function; if Then the current sending node is determined to be a valid node. The current sending node is deemed an illegal node, and the message is discarded.

[0031] Step 2.4: Repeat steps 2.1 to 2.3 until the next round of physical layer authentication is performed.

[0032] In step 1.2, and The maximum time reversal resonance intensity (MTRRS) and minimum Euclidean distance (SED) of the positive training samples are respectively, as shown in formulas (1) and (4);

[0033]

[0034] in,

[0035]

[0036]

[0037] Where, . represents the Frobenius norm, for Time reversal form;

[0038] express and The maximum value of the elements in the array generated after convolution;

[0039]

[0040] in,

[0041]

[0042] In step 1.3, and They are the maximum time reversal resonance intensity (MTRRS) and minimum Euclidean distance (SED) of the negative training samples, as shown in equations (7) and (8);

[0043]

[0044]

[0045] The beneficial effects of this invention are as follows:

[0046] 1. A physical layer authentication method for underwater acoustic networks that does not require training data from attackers is disclosed. The authentication training set can be obtained only through the interaction between legitimate neighbor nodes. This method makes full use of the spatial variation characteristics of underwater acoustic links and provides a new way to perform high-accuracy authentication in the absence of prior knowledge of illegal channels.

[0047] 2. Based on the use of multiple features from the CIR library, the time-varying patterns of the underwater acoustic link are jointly characterized, effectively overcoming the influence of the dynamically time-varying underwater acoustic channel, thereby further improving the accuracy of the physical layer authentication algorithm;

[0048] 3. The authentication method that combines a small sample underwater acoustic channel multi-feature training set with SVM algorithm for intelligent decision-making not only enables distributed spoofing attack detection at each node in the underwater acoustic network, but also significantly reduces the authentication overhead and network energy consumption in the underwater acoustic network. Attached Figure Description

[0049] Figure 1 This is the overall flowchart of the present invention.

[0050] Figure 2 This is a schematic diagram of the underwater acoustic network node layout of the present invention.

[0051] Figure 3 This is the distribution map of the multi-feature training data of the present invention. Figure 3 (a) is a distribution diagram of the training data for link AB1. Figure 3 (b) is a distribution diagram of the training data for link AB2.

[0052] Figure 4 This is a graph showing the certification performance results of the present invention. Figure 4 (a) is the receiver operating characteristic curve (ROC) plot for link AB1. Figure 4 (b) is the ROC curve of link AB2. Detailed Implementation

[0053] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0054] This invention comprises two phases: a network initialization phase and a transmission and authentication phase.

[0055] The main task of the network initialization phase is to construct a training set of channel features related to the CIR library in the absence of attacking nodes, and to train the optimal authentication decision boundary for the transmission and authentication phases. First, each legitimate receiving node in the underwater acoustic network establishes a CIR-based link library for each of its legitimate neighbors to overcome the influence of time-varying underwater acoustic channels. Second, time-resistant multi-features are extracted using the established CIR library, including maximum time-reversal resonance intensity (MTRRS) and minimum Euclidean distance (SED) based underwater acoustic channel features, to jointly characterize the time-varying patterns of underwater acoustic links. Next, based on the spatially varying characteristics of the underwater acoustic channel, positive and negative training samples are generated for each legitimate neighbor using the extracted multi-features to construct a training set in the absence of illegal channel information. Finally, to achieve accurate authentication while reducing authentication overhead and network energy consumption, the advantage of Support Vector Machine (SVM) algorithm with small sample sizes is utilized to train an SVM authentication model to analyze features and determine the optimal decision boundary for each legitimate neighbor.

[0056] The main task of the transmission and authentication phase is to use the optimal decision boundary trained for each legitimate neighbor during the network initialization phase to authenticate newly received messages. The aim is to effectively improve the authentication rate and the detection rate of attack nodes under time-varying underwater acoustic channels.

[0057] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0058] Figure 1 The overall flowchart of this invention is divided into two stages: network initialization and transmission and authentication.

[0059] Figure 2This diagram illustrates the deployment of nodes in an underwater acoustic network. The network includes a legitimate receiving node (Alice), legitimate transmitting nodes (Bob1 and Bob2), and an attacking node (Eve1). The initial positions (distance / meter, depth / meter) of each node are (0, 61.60), (1080, 85.75), (1080, 55.75), and (1080, 82.00), respectively. The CIR dataset used in this invention originates from the KAM11 sea trial conducted near the west coast of Kauai from June 23 to July 12, 2011.

[0060] Step 1: Network initialization phase;

[0061] The underwater acoustic network contains nodes Alice(0,61.60), Bob1(1080,85.75), and Bob2(1080,55.75). The implementation of this phase of the task includes the following four steps:

[0062] Step 1.1: Create the CIR library;

[0063] Alice received t m Moments from Bob i The probe signals (i = 1, 2) are used to estimate link AB using the copy correlation method. i At this moment In the same way, Alice estimated the link AB at different times. i The CIR, and thus the legal neighbor node Bob. i Establish a CIR link library M is link AB i The size of the CIR library;

[0064] Step 1.2: Construct positive training samples based on CIR library-related features;

[0065] To construct positive training samples for authenticating Bob1, Alice first utilizes the CIR library of link AB1. In The remaining M-1 CIRs calculate the features related to the two CIR libraries. and Next, Alice used these two features and Constitute positive training samples in, The process of constructing positive training samples for authenticating Bob2 is similar;

[0066] Step 1.3: Construct negative training samples based on CIR library-related features;

[0067] To construct negative training samples for authenticating Bob1, Alice first utilizes the CIR library of link AB2. In CIR library with link AB1 The M CIRs in the calculation of two features and Secondly, Alice utilizes features and Constructing negative training samples in, The process of constructing negative training samples for certified Bob2 is similar;

[0068] Step 1.4: Model training;

[0069] First, Alice uses the positive and negative training samples obtained in steps 1.2 and 1.3 to train Bob. i (i=1,2) Construct the training set in, This indicates that Bob is certified. i The q-th training sample is generated; then, Alice trains the SVM model to obtain Bob. i Optimal authentication boundary in, Bob i The normal vector of the optimal authentication boundary, b i Its bias term;

[0070] Step 2: Transmission and Authentication Phase;

[0071] Given nodes Alice(0,61.60), Bob1(1080,85.75), Bob2(1080,55.75), and Eve1(1080,82.00), this phase includes the following four steps:

[0072] Step 2.1: For a newly received message at time t', Alice obtains the IDi of the claimed sending node and estimates the current CIRh. i' ;

[0073] Step 2.2: Alice selects the corresponding node based on the sending node's IDi. Using CIRh t' and CIR library Calculate the features related to the two CIR libraries and And form the feature vector

[0074] Step 2.3: Alice uses the optimal authentication boundary obtained in Step 1.4 Calculate eigenvectors Authentication decision value like If the current sending node is deemed a valid node, it is deemed an invalid node, and the message is discarded.

[0075] Step 2.4: Repeat steps 2.1 to 2.3 until the next round of physical layer authentication is performed.

[0076] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

[0077] The underwater acoustic network physical layer authentication model constructed based on the method of this invention has been verified through sea trials, and the resulting multi-feature training data distribution is as follows: Figure 3 As shown, the results clearly demonstrate that the MTRRS value of legitimate underwater acoustic links is too high (>0.82), the SED value is too low (<7.39), and the authentication decision boundary can significantly distinguish between positive and negative training data. Under a small sample underwater acoustic channel feature training set, the method of this invention is compared with the time-reverse resonance-based underwater acoustic network node authentication (TRPLA) method proposed in ZL202111022616.5. The authentication performance is as follows: Figure 4 As shown, the ROC curves demonstrate the superiority of the method of this invention in improving the authentication rate. When the false alarm rate is 5%, Bob1's authentication rate increases to 94.5%, and Bob2's authentication rate reaches 100%, which reflects the high authentication accuracy of this invention under low-overhead conditions.

Claims

1. A method for multi-feature intelligent physical layer authentication of underwater acoustic networks, characterized in that Comprising the following steps: Step 1: Network initialization phase; A legal receiving node Alice, whose legal neighbor node is Bob i , 1≤i≤N, N is the number of legal neighbor node Bob i , all legal nodes are anchored at different depths of the ocean and will drift with the ocean current; the underwater acoustic link between node Alice and node Bob i is denoted as AB i ; Step 1.1: Create CIR library; In t m At that moment, Alice received a message from Bob. i The probe signal, 1≤i≤N, is used to estimate the link AB using the copy correlation method. i At this moment in, For t m Time Link AB i The complex amplitude of the l-th path of the channel impulse response (CIR), 0 ≤ l ≤ L-1, where L is the total number of channel paths; to cope with the influence of time-varying underwater acoustic channels, Alice sets up a complex amplitude for each of its legitimate neighbor nodes Bob. i Each one establishes a CIR library 1≤i≤N, 1≤m≤M, where M is link AB i The size of the CIR library, Record link AB i CIR sequences at different times were used to capture patterns of CIR variation over time. Step 1.2: Construct positive training samples based on the relevant features of the CIR library; In each legal neighbor node Bob i After establishing the CIR library, in order to construct information about the legal neighbor node Bob i First, Alice uses the link AB to obtain positive training samples. i CIR library any one of them Together with the remaining M-1 CIRs, calculate the multi-features related to the CIR library. F is the number of features, 1≤i≤N. 1≤s≤M; Next, Alice uses the CIR library related multi-features constitute positive training samples That is: wherein, is a label of the training data (feature vector) ; Step 1.3: Construct negative training samples based on the relevant features of the CIR library; First, Alice uses link AB j CIR library In and link AB i CIR library Given M CIRs, 1≤s≤M, 1≤i≤N, j≠i, calculate the multi-features related to the CIR library. F is the number of features; Next, Alice uses the CIR library related multi-features constitute negative training samples That is: wherein, is a label of the training data (feature vector) ; Step 1.4: Model training; First, Alice uses the positive training samples and negative training samples obtained in steps 1.2 and 1.3 to train a support vector machine (SVM) model for her own legitimate neighbor node Bob i Constructing the training set 1≤i≤N, where, Bob is authenticated i The qth training sample is constructed; then, Alice inputs the training set into a support vector machine (SVM) model for training, thereby obtaining the optimal authentication boundary of Bob i That is:​ wherein, represents the normal vector of the maximum authentication boundary authenticating Bob i , represents the Lagrange multiplier of the qth training sample authenticating Bob i , φ(.) represents a kernel function, b i represents the bias term of the maximum authentication boundary authenticating Bob i , represents the vth unbounded support vector authenticating Bob i , 0 < v < |V i |, V i represents the set of unbounded support vectors authenticating Bob i . Step 2: Transmission and authentication phase; A legal receiving node Alice, a legal neighbor node Bob of Alice i , 1≤i≤N, N is the number of legal neighbor nodes Bob i , 1≤j≤K, K is the number of attack nodes; attack node Eve j , 1≤j≤K, K is the number of attack nodes; attack node Eve j uses the same node identity ID as Bob i Send data packets, aiming to obtain network access rights, so as to further launch a series of malicious attacks and destroy the normal operation of the underwater acoustic network; Step 2.1 : For a newly received message at time t', Alice obtains the ID of the sending node as claimed by the message i and estimates the current Step 2.2: Alice selects the corresponding i Based on the formula (7) and (8), the CIRh t' and CIR library Calculate the multi-features related to CIR library and form the feature vector ​ Step 2.3: Alice uses the most authenticated boundary obtained in step 1.4 Computing the feature vector The authentication decision value where sgn(.) is the sign function and φ(.) is the kernel function; if decides that the current sending node is a legal node, if decides that the current sending node is an illegal node and discards the message; Step 2.4: Repeat steps 2.1 to 2.3 until the next round of physical layer authentication is performed.

2. The underwater acoustic network multi-feature intelligent physical layer authentication method according to claim 1, characterized in that: In step 1.2, and The maximum time reversal resonance intensity (MTRRS) and minimum Euclidean distance (SED) of the positive training samples are respectively, as shown in formulas (1) and (4); Wherein, where ||. || denotes the Frobenius norm, is the time-reversed form of is the time-reversed form of representing and the maximum value of the elements of the array generated after convolution; Wherein, 3. The underwater acoustic network multi-feature intelligent physical layer authentication method according to claim 1, characterized in that: The step 1.3 in, and are the negative training sample maximum time-reversed resonance strength MTRRS and minimum Euclidean distance SED, respectively, as shown in equations (7) and (8);

Images