Low-altitude wireless network high-precision interference attack detection method based on unmanned aerial vehicle group

Abstract

The invention discloses a low-altitude wireless network high-precision interference attack detection method based on an unmanned aerial vehicle group, and belongs to the technical field of wireless communication, network security and signal detection. According to the method, a plurality of unmanned aerial vehicles are used for measuring communication performances such as received signal strength, channel state, network feedback transmission delay and packet loss rate of a low-altitude wireless network, and hypothesis testing is constructed based on multivariate statistical characteristics to detect interference attacks such as static interference, reactive interference, random interference, sweep frequency interference, comb interference and intelligent interference. The method is suitable for smart cities, low-altitude economy, military communication, emergency communication and other scenes. The unmanned aerial vehicle group sharing detection experience can jointly mine the network interference attack situation, and reinforcement learning is adopted to continuously optimize the detection threshold, so that the false alarm rate and the missing report rate are reduced while the detection precision is ensured. The comprehensive monitoring capability, the refined classification capability, the efficient cooperation mechanism, the intelligent optimization capability and the high-precision detection performance are shown in a low-altitude wireless network scene.

Description

Technical Field

[0001] The present invention belongs to the technical fields of wireless communication, network security and signal detection, and specifically relates to a high-precision interference attack detection method for low-altitude wireless networks based on drone swarms. Background Art

[0002] With the continuous evolution of wireless communication technology, interference attacks have become one of the major challenges facing wireless communication systems. Interference detection technology aims to provide key information such as interference type and intensity to operation and maintenance personnel by monitoring the characteristics of wireless signals, guiding the effective implementation of interference countermeasures and ensuring the quality of wireless communication network services.

[0003] Interference detection is usually considered as a hypothesis testing problem, and the presence of interference in a wireless network can be determined by comparing the test statistic with the test threshold. The literature [L.Arcangeloni, E.Testi and A.Giorgetti, "Detection of Jamming Attacks via Source Separation and Causal Inference," in IEEE Transactions on Communications, vol.71, no.8, pp.4793-4806, Aug.2023] considers that there is a causal relationship between the reactive jammer with monitoring behavior and the legitimate user transmission, and uses the transfer entropy as the test statistic, and compares it with the preset threshold to achieve interference detection. The Chinese patent application with application publication number CN117169923A proposes an interference detection method for a global navigation satellite system, which uses satellite signals, ephemeris data, satellite positions, satellite speeds and other information, and uses a weighted method to calculate the discrimination coefficient, and compares it with the preset threshold to achieve interference detection. However, due to the complexity and variability of the environment, it is difficult to obtain prior information on attack patterns in a timely manner.

[0004] The data-driven interference detection method uses observed data to learn the characteristics of interference signals and realizes the detection of various types of interference attacks. Reference [S.Sciancalepore,F.Kusters,NKAbdelhadi et al.,"JammingDetection in Low-BERMobileIndoorScenarios via Deep Learning,"in IEEEInternet of Things Journal,vol.11,no.8,pp.14682-14697,Apr.,2024] Considering that interference attacks will cause CSI anomalies, an autoencoder is used to process the CSI matrix to realize interference detection. Reference [C.Wang,Y.Chen,Z.Linet al.,"Reinforcement learning based jamming detection forreliable wirelesscommunications,"in Proc.IEEE Veh.Tech.Conf.,Singapore,June 2024] uses a deep Q-learning algorithm to select the test threshold, evaluate the channel energy and packet loss rate of the receiver, and improve the detection accuracy. Reference [Gong Xiaoxue, Pang Jiahao, Zhang Qihan, et al. Optical network interference attack detection, identification and recovery method based on machine learning [J]. Journal of Communications, 2023, 44(07): 159-170] proposed a bidirectional long short-term memory detection model to automatically mine and process the correlation of state information between spectral data to perform attack detection and identification. The Chinese patent application publication number CN117394934A uses a multi-branch convolutional neural network as an interference identification network. By training the neural network, the network automatically extracts the characteristics of the interference signal and iteratively updates all trainable parameters through the back propagation algorithm to improve the accuracy of the interference identification model.

[0005] The change in the position of the transmitting terminal causes the strength of the received signal to increase and decrease. Although it is legal, it is easy to be confused with the behavior of the jammer, which in turn affects the accuracy of the jammer detection. The monitoring range of drones for low-altitude wireless networks is wide, and the legal signal fluctuates greatly, which makes it more difficult to distinguish between real signal changes and jammer behavior. Summary of the invention

[0006] The purpose of the present invention is to provide a high-precision interference attack detection method for low-altitude wireless networks based on drone swarms, which uses drone swarms to measure multiple features such as received signal strength in low-altitude wireless networks, thereby constructing hypothesis tests to identify interference attacks; in addition, detection experience is shared among multiple drones, and reinforcement learning methods are used to continuously optimize detection thresholds to improve interference detection accuracy.

[0007] The method for detecting high-precision interference attacks on low-altitude wireless networks based on drone swarms includes the following steps:

[0008] Step 1: Initialize the drone queue: Each drone maintains two queues W1 and W2. The length of queue W1 is L, which is used to record the number of false alarms in historical detection results. F and the number of missed reports l M ; The length of queue W2 is N, which is used to count the historical interference-free feature set; define the detection threshold set X∈{r / R|0≤r≤R}, where the maximum optional threshold value is 1, and the total is quantized into R levels; the drone swarm performs K rounds of feature sampling to obtain wireless network features; each round of sampling time is one time slot, and the total time slots are {1,2,L,k,L,K};

[0009] Step 2: Network initialization: Construct the benefit network Q and risk network E, both of which are four-layer fully connected neural networks, for learning and predicting detection benefits and risks, and initialize network parameters, including the number of neurons in each layer, learning rate, and experience pool;

[0010] Step 3: The wireless network characteristic measurement and statistics process includes measuring the received signal strength, receiving low-altitude network broadcast messages, calculating the false alarm rate and missed alarm rate, and interacting with neighbors;

[0011] Step 4: Construct the state vector s (k) =[m,g,δ,ρ,η,χ,{f j} 1≤j≤M ,{g j} 1≤j≤M ];

[0012] Step 5: Select the detection threshold x according to the following strategy (k) :

[0013]

[0014] Where Pr(·) is the threshold selection probability, x * represents the optimal detection threshold, Q(s (k) ,x) and E(s (k) ,x) represent the reward value and risk value corresponding to the selected detection threshold respectively;

[0015] Step 6: Calculate the average received signal strength and variance The average transmission delay and variance The average value of packet loss rate η and variance The average value of the bit error rate χ and variance Calculate the test statistic Δ based on the feature weights w1, w2, and w3:

[0016]

[0017] Where ξ is a binary indicator symbol, which is represented by ξ=1 when the drone successfully decodes the receiver broadcast message of the low-altitude wireless network, otherwise ξ=0;

[0018] Step 7: Compare the test statistic Δ with the detection threshold x. If Δ≤x, it means no interference signal is detected. Update the no-interference detection result to ψ i =0; otherwise, broadcast interference exists alarm ψ i =1;

[0019] Step 8: Receive the signal that satisfies the false alarm rate f j ≤γ1, false negative rate g j ≤γ2 condition The detection results shared by neighboring drones, where γ11∈[0,1] and γ2∈[0,1] represent the historical accuracy requirements of neighboring drones participating in the decision-making, ψ i ′ is represented by the local result ψ i The complement of the strategy; calculate the risk value e (k) =w s I(Σψ i ′-λ≥0|ψ=0), where the risk weight for The historical detection accuracy expectation of the UAV; I(g) is the indicator function, when the conditions in the brackets are met, I(g) = 1, otherwise I(g) = 0, λ represents the number of votes required to reach a group consensus, ψ = 0 means that the local UAV is judged to be non-interfering;

[0020] Step 9: Use the received signal strength δ (k) , average transmission delay ρ (k) , packet loss rate η (k) and bit error rate χ (k) Update queue W2, verify the test results and update queue W1;

[0021] Step 10: Calculate the detection benefit u (k) =-c F f F -c M f M , where the weight c F and c M Respectively represent the importance attached to false alarm rate and missed detection rate;

[0022] Step 11: Generate experience h (k) =[s (k) ,x (k) ,u (k), e (k) , s (k+1) ] and store it in the experience pool D; where s (k) , x (k) , u (k) , e (k) are the state, selection detection threshold, detection benefit, and strategy risk value of the previous time slot respectively, and s (k+1) is the state of the next time slot;

[0023] Step 12: When the number of stored experiences is more than B, randomly select B experiences from the experience pool D, and update the weight parameters θ and φ using the stochastic gradient descent algorithm and the backpropagation algorithm of the neural network as shown in the following formula:

[0024]

[0025] where α and β are the learning factors of the reward value and the risk value respectively;

[0026] Step 13: Repeat steps 3 - 12 until |Q(s (k+1) , x (k+1) ) - Q(s (k) , x (k) )| < δ and |E(s (k+1) , x (k+1) ) - E(s (k) , x (k) )| < ε, and the algorithm converges; where δ and ε are the convergence judgment conditions of the reward value and the risk value respectively.

[0027] In step 1, the wireless network characteristics include the received signal strength measured at the drone, and information such as the channel state, average transmission delay, packet loss rate, and bit error rate between the transceiver feedback by the low - altitude wireless network.

[0028] In step 2, both the benefit network Q and the risk network E include four fully - connected layers. The input layer contains 2M + 6 neurons; the second and third layers are hidden layers, containing H and G neurons respectively; the output layer contains R neurons; initialize θ ∈ [0, 1], φ ∈ [0, 1], α ∈ [0, 1], β ∈ [0, 1], λ ∈ [0, M], γ1 ∈ [0, 1], γ2 ∈ [0, 1], w1 ∈ (0, 1), w2 ∈ (0, 1), w3 ∈ (0, 1), c F ∈ (0, 1), c M ∈ (0, 1), δ ∈ (0, 1), ε ∈ (0, 1), the initial experience pool B ∈ (0, 128), the initial detection result ψ (0) = 0.

[0029] In step 3, the specific steps of wireless network feature measurement and statistics include: measuring the received signal strength δ at the location of the receiver in the k time slot; receiving a message broadcast by a receiver such as a base station in a low-altitude wireless network, including the channel state g, average transmission delay ρ, packet loss rate η and bit error rate χ between the receiver and the transmitter; calculating the false alarm rate f i = l F / L and false alarm rate g i = l M / L, where l F Indicates the number of false alarms, l M represents the number of missed alarms, L represents the queue length; it exchanges beacons with neighbors to obtain the number of surrounding drones m, and receives their broadcast messages to count the false alarm rate of the drone group {f j} 1≤j≤M Missing rate {g j} 1≤j≤M .

[0030] Compared with the prior art, the present invention has the following outstanding advantages:

[0031] 1. For low-altitude wireless network scenarios, multiple drones are used to measure the communication performance of low-altitude wireless networks, such as the received signal strength, channel status, and transmission delay and packet loss rate of network feedback. The distributed monitoring strategy effectively avoids the blind spot problem that may exist in single-point monitoring and ensures the comprehensiveness and accuracy of the data.

[0032] 2. By constructing hypothesis tests based on multivariate statistical features, different types of interference attacks can be accurately distinguished, including static interference, reactive interference, random interference, sweep interference, comb interference, and intelligent interference. Refined classification capabilities help to take targeted defense measures; the use of multivariate statistical features can more accurately capture the characteristics of interference signals, thereby improving the accuracy and reliability of detection;

[0033] 3. The shared detection experience of drone groups can jointly mine the network interference attack situation, and use reinforcement learning to continuously optimize the detection threshold and improve the detection accuracy. By integrating the monitoring data of multiple drones, it is possible to have a more comprehensive understanding of the network status, timely detect potential interference threats, and improve the adaptability and robustness of the system.

[0034] 4. The reinforcement learning algorithm is used to optimize the interference hypothesis detection threshold. By continuously learning and adapting to changes in the network environment, the system can automatically adjust the detection parameters to cope with interference attacks in different scenarios, improve the interference detection accuracy and robustness in intelligent interference scenarios with power control, and reduce the false alarm rate and missed alarm rate while ensuring the detection accuracy.

[0035] 5. The present invention demonstrates comprehensive monitoring capabilities, refined classification capabilities, efficient collaboration mechanisms, intelligent optimization capabilities and high-precision detection performance when facing low-altitude wireless network scenarios, providing strong guarantees for the security protection of low-altitude wireless networks. The present invention is suitable for scenarios such as smart cities, low-altitude economy, military communications, and emergency communications. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is the false alarm rate of the high-precision interference detection method described in the embodiment of the present invention.

[0037] Figure 2 It is the false alarm rate of the high-precision interference detection method described in the embodiment of the present invention.

[0038] Figure 3 This is the benefit of the high-precision interference detection method described in the embodiment of the present invention. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the following embodiments will further illustrate the present invention in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0040] The embodiment of the present invention includes the following steps:

[0041] Step 1: A total of M drones jointly observe low-altitude wireless networks. Each drone initializes a queue W1 of length L to record historical detection results, where the number of false alarms and missed alarms are represented by l and l respectively. F , l M , initialize the queue W2 of length N to count the historical interference-free feature set. The drone swarm performs K rounds of feature sampling, each round of sampling time is one time slot, and the total time slot is {1,2,L,k,L,K}. After each sampling, the drone can select R detection thresholds, which are predefined as the set X={r / R|0≤r≤R}; construct the benefit network Q and risk network E, both of which contain four fully connected layers, of which the input layer contains 2M+6 neurons; the second and third layers are hidden layers, containing H and G neurons respectively; the output layer contains R neurons, and R neurons correspond to the evaluation results of R detection thresholds. Initialize the Q network weight parameter θ∈[0,1], the E network weight parameter φ∈[0,1], the Q network discount factor α∈[0,1], the E network discount factor β∈[0,1], the number of votes required to reach group consensus λ∈[0,M], the historical detection accuracy requirements γ1∈[0,1], γ2∈[0,1], the feature weights w1∈(0,1), w2∈(0,1), w3∈(0,1), the false alarm rate coefficient c F ∈(0,1), missed detection rate coefficient c M∈(0,1), detection termination conditions ò∈(0,1), ε∈(0,1), initial experience pool (empty set), set the experience pool size B∈(0,128), the initial detection result ψ (0) =0 (indicating the initial state without interference).

[0042] In this embodiment, the number of drones is set to M = 5, the total number of sampling times is K = 2000, and the single sampling time is 10s. The maximum length of the queue W1 is set to L = 100, where the number of false alarms and missed detections is counted as l F and l M , set the maximum length of queue W2 N = 30, and the detection threshold set X = {r / 20|0≤r≤20}. Construct the benefit network Q and risk network E, and the number of neurons in the four fully connected layers is 16, 64, 128 and 20 respectively. Initialize θ = φ = 0, α = 0.01, β = 0.05, λ = 3, γ1 = γ2 = 0.25, B=32,ψ (0) =0, w1=w2=w3=0.3, c F =c M =0.5,ò=ε=0.01.

[0043] Step 2: In time slot k, measure the received signal strength δ at the location where the receiver is located; receive the message broadcast by the receiver such as the base station in the low-altitude wireless network, including the channel state g, average transmission delay ρ, packet loss rate η and bit error rate χ between the receiver and the transmitter; calculate the false alarm rate f i = l F / L and false alarm rate g i = l M / L, where l F Indicates the number of false alarms, l M represents the number of missed alarms, L represents the current length of queue W1; the time division multiple access TDMA communication mechanism is used to interact with neighbor beacons to obtain the number of surrounding drones m, and receive their broadcast messages to count the false alarm rate of the drone group {f j} 1≤j≤M and the false negative rate {g j} 1≤j≤M .

[0044] Step 3: Construct the state vector s (k) =[m,g,δ,ρ,η,χ,{f j} 1≤j≤M ,{g j} 1≤j≤M ].

[0045] Step 4: Select the detection threshold x according to the following formula (k) :

[0046]

[0047] Among them, Pr(·) is the action selection probability, x * represents the optimal detection threshold, Q(s (k) ,x) and E(s (k) ,x) represent the reward value and risk value corresponding to the selected detection threshold respectively.

[0048] Step 5: Calculate the average received signal strength and variance The average transmission delay and variance The average value of packet loss rate η and variance The average value of the bit error rate χ and variance Calculate the test statistic Δ based on the feature weights w1, w2, and w3:

[0049]

[0050] Among them, ξ is a binary indicator symbol, which is represented by ξ=1 when the drone successfully decodes the receiver broadcast message of the low-altitude wireless network, otherwise ξ=0.

[0051] Step 6: Compare the test statistic Δ with the detection threshold x. If Δ≤x, it means no interference signal is detected. Update the no interference detection result to ψ i =0; otherwise, broadcast interference exists alarm ψ i =1.

[0052] Step 7: Receive the contention f j ≤γ1,g j ≤γ2 condition The detection results shared by neighboring drones are represented as local results ψ i The complement of i '; Calculate the risk value of the strategy e (k) =w s I(Σψ i ′-λ≥0|ψ=0), where the risk weight for The historical detection accuracy expectation of the UAV. Among them, I(g) is the indicator function, when the conditions in the brackets are met, I(g) = 1, otherwise I(g) = 0, λ represents the number of votes required to reach a group consensus, and ψ = 0 means that the local UAV is judged to be non-interfering.

[0053] Step 8: Use the received signal strength δ (k) , average transmission delay ρ (k) , packet loss rate η (k) and bit error rate χ (k)Update queue W2 and verify the detection result to update queue W1.

[0054] Step 9: Calculate u (k) =-c F f F -c M f M . Among them, the weights c F and c M respectively represent the degrees of emphasis on the false alarm rate and the missed detection rate.

[0055] Step 10: Generate experience h (k) =[s (k) , x (k) , u (k) , e (k) , s (k+1) and store it in the experience pool D. Among them, s (k) , x (k) , u (k) , e (k) are respectively the state, the selected detection threshold, the detection benefit, and the policy risk value of the previous time slot, and s (k+1) is the state of the next time slot.

[0056] Step 11: When the number of stored experiences is more than B, randomly select B experiences from the experience pool D, and update the weight parameters θ and φ using the stochastic gradient descent algorithm and the backpropagation algorithm of the neural network as shown in the following formula:

[0057]

[0058] Among them, α and β are the Q-network discount factor and the E-network discount factor respectively.

[0059] Step 12: Repeat steps 2 to 11 until |Q(s (k+1) , x (k+1) ) - Q(s (k) , x (k) )| < ò and |E(s (k+1) , x (k+1) ) - E(s (k) , x (k) )| < ε, and the algorithm converges; where ò and ε are the convergence judgment conditions for the reward value and the risk value respectively, ò ∈ (0, 1), ε ∈ (0, 1).

[0060] To verify the effect of the embodiments of the present invention, in a laboratory scenario, a USRP N210 is used to transmit an interference signal with a maximum of 100 mW to interfere with the video transmission based on a Raspberry Pi, realizing the acquisition of interference characteristics including the received signal strength, transmission experiment, packet loss rate, bit error rate, etc., and performing interference detection simulation based on the collected characteristics, and showing the chart analysis of three key indicators. Figure 1 is the false alarm rate of the high-precision interference detection method according to the embodiment of the present invention, Figure 1 It can be seen that the method of the present invention significantly reduces the false alarm rate, indicating that the system can more accurately determine when no interference occurs, reducing unnecessary alarms and erroneous operations. Figure 2 is the underreporting rate of the high-precision interference detection method described in the embodiment of the present invention. Figure 2 It can be seen that the method of the present invention effectively reduces the false alarm rate and enhances the system's sensitivity and detection capability to real interference, which is crucial to ensuring the stability and reliability of the communication system. Figure 3 This is the benefit of the high-precision interference detection method described in the embodiment of the present invention. Figure 3 It can be seen that the method of the present invention significantly improves the detection efficiency, that is, the system can more accurately identify and respond to interference events, thereby improving the performance of the overall communication system and user experience. In summary, the embodiments of the present invention have significant advantages in reducing false alarm rate, missed alarm rate and improving detection efficiency. It is of great significance to improve the anti-interference ability and stability of video transmission systems based on low-cost hardware platforms such as Raspberry Pi, and is expected to play an important role in practical applications.

[0061] The present invention uses multiple drones to measure the received signal strength, channel status, transmission delay and packet loss rate of low-altitude wireless networks, and constructs hypothesis tests based on multivariate statistical features to detect interference attacks such as static interference, reactive interference, random interference, sweep interference, comb interference and intelligent interference. It is suitable for scenarios such as smart cities, low-altitude economy, military communications, and emergency communications. Reinforcement learning is used to continuously optimize the detection threshold, and the drone group shares the detection experience to jointly explore the network interference attack situation and improve the detection accuracy.

[0062] The above embodiments are only preferred embodiments of the present invention and cannot be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A high-precision jamming attack detection method for low-altitude wireless networks based on drone swarms, characterized by The following steps are involved: Step 1: Initialize the drone queue: Each drone maintains two queues W1 and W2. The length of queue W1 is L, which is used to record the number of false alarms in historical detection results. F and the number of missed reports l M ; The length of queue W2 is N, which is used to count the historical interference-free feature set; define the detection threshold set X∈{r / R|0≤r≤R}, where the maximum optional threshold value is 1, and the total quantization is R levels; The drone swarm performs K rounds of feature sampling to obtain wireless network features. Each round of sampling lasts for one time slot, and the total time slots are {1, 2, L, k, L, K}. Step 2: Network initialization: Construct the benefit network Q and risk network E, both of which are four-layer fully connected neural networks, for learning and predicting detection benefits and risks, and initialize network parameters, including the number of neurons in each layer, learning rate, and experience pool; Step 3: The wireless network characteristic measurement and statistics process includes measuring the received signal strength, receiving low-altitude network broadcast messages, calculating the false alarm rate and missed alarm rate, and interacting with neighbors; Step 4: Construct the state vector s (k) =[m,g,δ,ρ,η,χ,{f j } 1≤j≤M ,{g j } 1≤j≤M ]; Step 5: Select the detection threshold x according to the following strategy (k) : Where Pr(·) is the threshold selection probability, x * represents the optimal detection threshold, Q(s (k) ,x) and E(s (k) ,x) represent the reward value and risk value corresponding to the selected detection threshold respectively; Step 6: Calculate the average received signal strength and variance The average transmission delay and variance The average value of packet loss rate η and variance The average value of the bit error rate χ and variance Calculate the test statistic Δ based on the feature weights w1, w2, and w3: Where ξ is a binary indicator symbol, which is represented by ξ=1 when the drone successfully decodes the receiver broadcast message of the low-altitude wireless network, otherwise ξ=0; Step 7: Compare the test statistic Δ with the detection threshold x. If Δ≤x, it means no interference signal is detected. Update the no-interference detection result to ψ i =0; otherwise, broadcast interference exists alarm ψ i =1; Step 8: Receive the signal that satisfies the false alarm rate f j ≤γ1, false negative rate g j ≤γ2 condition The detection results shared by neighboring drones are: γ1∈[0,1] and γ2∈[0,1] represent the historical accuracy requirements of neighboring drones participating in the decision-making, and the strategy risk value e is calculated. (k) =w s I(∑ψ i ′-λ≥0|ψ=0), where the risk weight for The historical detection accuracy expectation of the UAV; I(g) is the indicator function, when the conditions in the brackets are met, I(g) = 1, otherwise I(g) = 0; ψ i ′ is represented by the local result ψ i The complement of ; λ represents the number of votes required to reach a group consensus, and ψ=0 means that the local drone judges that there is no interference; Step 9: Use the received signal strength δ (k) , average transmission delay ρ (k) , packet loss rate η (k) and bit error rate χ (k) Update queue W2, verify the test results and update queue W1; Step 10: Calculate the detection benefit u (k) =-c F f F -c M f M , where the weight c F and c M Respectively represent the importance attached to false alarm rate and missed detection rate; Step 11: Generate experience h (k) =[s (k) ,x (k) ,u (k) ,e (k) ,s (k+1) ] and stores it in the experience pool D; where s (k) 、x (k) 、u (k) 、e (k) are the state of the previous time slot, the selected detection threshold, the detection benefit, and the strategy risk value, respectively. (k+1) is the state of the next time slot; Step 12: When the number of stored experiences is greater than B, randomly select B experiences from the experience pool D, and use the stochastic gradient descent algorithm and the back propagation algorithm of the neural network to update the weight parameters θ and φ, as shown in the following formula: Among them, α and β are learning factors of reward value and risk value respectively; Step 13: Repeat Steps 3 to 12 until |Q(s (k+1) , x (k+1) ) - Q(s (k) , x (k) )| < δ and |E(s (k+1) , x (k +1) ) - E(s (k) , x (k) )| < ε, at which point the algorithm converges; where δ and ε are the convergence judgment conditions for the reward value and the risk value, respectively.

2. The method for detecting low-altitude wireless network interference attacks based on drone swarms as claimed in claim 1 is characterized in that In step 1, the wireless network characteristics include the received signal strength measured at the drone, and the channel status, average transmission delay, packet loss rate and bit error rate between transceivers fed back by the low-altitude wireless network.

3. The method for detecting low-altitude wireless network interference attacks based on drone swarms as claimed in claim 1 is characterized in that In step 2, the benefit network Q and the risk network E both contain four fully connected layers, where the input layer contains 2M+6 neurons; the second and third layers are hidden layers, containing H and G neurons respectively; the output layer contains R neurons; initialize θ∈[0,1], φ∈[0,1], α∈[0,1], β∈[0,1], λ∈[0,M], γ1∈[0,1], γ2∈[0,1], w1∈(0,1), w2∈(0,1), w3∈(0,1), c F ∈(0,1), c M ∈(0,1), ò∈(0,1), ε∈(0,1), initial experience pool B∈(0,128), initial detection result ψ (0) =0.

4. The method for detecting low-altitude wireless network interference attacks based on drone swarms as claimed in claim 1 is characterized in that In step 3, the specific steps of wireless network feature measurement and statistics include: measuring the received signal strength δ at the location of the receiver in the k time slot; receiving a message broadcast by a receiver such as a base station in a low-altitude wireless network, including the channel state g, average transmission delay ρ, packet loss rate η and bit error rate χ between the receiver and the transmitter; calculating the false alarm rate f i = l F / L and false alarm rate g i = l M / L, where l F Indicates the number of false alarms, l M represents the number of missed alarms, L represents the queue length; it exchanges beacons with neighbors to obtain the number of surrounding drones m, and receives their broadcast messages to count the false alarm rate of the drone group {f j } 1≤j≤M Missing rate {g j } 1≤j≤M .

Images