Explanatable slice-level active threat prediction method in crowd sensing

Abstract

A large amount of high-value data contained in a mobile crowd sensing platform becomes targets of various threats such as Sybil attack and Dos attack, the sensing service of the platform is seriously interfered, and the sensing data quality is influenced. According to the learning and interpretable slice-level active threat prediction method provided by the invention, accurate long-term prediction of threats is provided, so that a sensing platform can formulate more appropriate defense means and strategies in advance. The method comprises the following steps: firstly, learning the statistical attribute of each slice through a'division and treatment 'non-stationary factor processing method, and predicting the future distribution of each slice; then, a similarity coefficient between the attack data and the time sequence base is learned in a comparative learning mode in combination with a bidirectional cross attention mechanism, and the stationary output of the stage is obtained by combining the similarity coefficient and the future part of the time sequence base; and finally, according to the prediction of the statistical attribute of the threat slice, recovering the non-stationary characteristic in the stationary output of the threat prediction module, and obtaining a final prediction result of the threat.

Description

Technical Field

[0001] The present invention belongs to the field of mobile crowd intelligence perception, and specifically relates to an explainable slice-level active threat prediction method. Background Art

[0002] Mobile Crowdsensing (MCS) uses smart mobile devices with diverse sensing functions. Relying on the ubiquity and connectivity of these smart devices, it is completely changing the way data is collected, processed, and analyzed for environmental information. The application scope of MCS covers many fields such as environmental monitoring, traffic management, natural disaster monitoring, and medical services. The data collected by MCS usually contains a large amount of personal information, location data, and environmental information. These data are very attractive to malicious attackers, making MCS face diverse, complex, and continuous serious security threats. Security threats are inevitable. Various attacks such as witch attacks, malicious attacks, Dos attacks, and phishing have posed a major challenge to the security and credibility of MCS.

[0003] Threats are inevitable. Various attacks such as witch attacks, malicious attacks, DoS attacks, phishing, etc. pose significant challenges to the security and credibility of MCS. Existing MCS defenses are still reactive in nature, involve a lot of overhead in terms of execution time, and are easily bypassed by attackers. This time delay is caused by a variety of reasons, such as the complex pattern matching required to identify the signatures of polymorphic malware, the complexity of attack data processing, or the complexity of security policies. Recently, machine learning-based anomaly detection algorithm models have been introduced. Although these models show good detection capabilities for unknown attack behaviors and are expected to automatically perform the tasks of detecting, tracking, and blocking malware and intruders. However, they always lag behind the occurrence of attacks, and there is a tendency to identify benign behaviors as abnormal behaviors, which causes a lot of resource waste to the platform, interferes with perception activities, and reduces the credibility of the platform.

[0004] The only research on attack prediction focuses on short-term predictions of hours or days, and the platform does not have enough reaction time to build defense strategies and tools. Today, long-term predictions of attacks are mainly based on the subjective perception of human experts, which is compromised by the lack of security knowledge. In addition, since security issues can directly lead to serious consequences such as the leakage of MCS perception data, privacy infringement, and task tampering, the system has much higher credibility requirements for the methods used than in other fields, which is often ignored by existing methods.

[0005] Compared with models that can be explained after the fact, models that can be explained before the fact provide a direct understanding of the internal logic and decision-making process of the model, which enhances the transparency and credibility of the model and makes it easier for users to accept and trust the model. This makes the ex ante explainability of the method one of the research focuses. Data availability has been proven to be effective in achieving proactive defense, taking action before potential threats escalate into actual events, and taking a proactive approach to mitigate, delay, or even prevent attacks. The threats faced by MCS are no exception. Deep learning models can provide accurate and reliable long-term predictions, take more proactive security operation methods, conduct low-latency defense, formulate appropriate countermeasures in advance, reveal security vulnerabilities and system weaknesses, and support decision makers in formulating more appropriate defense measures and strategies, such as Figure 1 shown. Summary of the invention

[0006] In response to the above problems, the present invention starts from three perspectives: non-stationary factor processing of threat data, threat prediction, and ex ante interpretability of the model, and realizes active long-term prediction of MCS threats through the designed model. The non-stationary prediction task is decoupled into a statistical prediction task and a stable time series prediction task. The goal of the statistical prediction task is to optimize the distribution similarity between the non-normalized output and the attack data, and the goal of the stable time series prediction task is to improve the accuracy of threat prediction. The overall framework is as follows: Figure 2 As shown, it mainly includes the following three parts:

[0007] (1) Non-stationary factor processing. A “divide and conquer” non-stationary factor processing method is proposed. The attack data is normalized by slice, and the statistical properties of each slice are learned. The statistical properties of future prediction data are obtained through the designed statistical prediction module. Based on the prediction of the statistical properties of the data by the statistical prediction module, its non-stationary characteristics are restored in the stable output of the threat prediction module to obtain the final prediction result of the security threat.

[0008] (2) Threat prediction. The data without non-stationary factors is used as the standardized input of the threat prediction module. The similarity coefficient between the attack data and the time series basis is learned by contrastive learning combined with the bidirectional cross-attention mechanism. The similarity coefficient and the future part of the time series basis are combined to obtain the stable output of this stage. An accurate security threat prediction method is proposed, which can adaptively and self-supervise the learning of the time series basis. The use of contrastive learning ensures the consistency of the time series basis in both the historical and future views. The future component of the time series basis is selected and consolidated according to the calculated similarity coefficient, thereby generating accurate future predictions.

[0009] (3) Model interpretability. The low ex ante interpretability of the MCS security threat prediction model will greatly reduce the credibility of the model and directly affect the priority and efficiency of security incident handling. In the non-stationary factor processing stage, the designed statistical prediction module is a linear model with good interpretability. The core of the threat prediction module is a learnable and interpretable time series basis. The time series basis is not set by humans, but is learned autonomously by the model based on the specific pattern of the data.

[0010] Compared with the prior art, the beneficial effects of the present invention are: (1) The non-stationary properties and evolution trends of threat data in the MCS are independently modeled from the perspective of slices, and the future distribution of non-stationary factors is estimated through explicit learning. Compared with the existing data processing from the overall perspective, the localized approach can better retain the data pattern of each slice itself, ensure the stability of the data change pattern over time, and achieve high-precision long-term prediction tasks. (2) From the perspective of a learnable and interpretable time series basis, the comparative learning method is combined with the bidirectional cross-attention mechanism to ensure the consistency of the time series basis in both the historical and future views. This precise threat prediction method further improves the accuracy of threat prediction. (3) Based on the need for high credibility of the method due to security issues, the present invention pays special attention to the ex ante interpretability of the model, and designs an interpretable statistical prediction module and a learnable and interpretable time series basis. Compared with the existing methods, the understanding and trust in the model prediction results are enhanced. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is the MCS threat prediction scenario diagram;

[0012] Figure 2 It is an interpretable threat prediction model graph. DETAILED DESCRIPTION

[0013] The historical attack data of threats and the key events that occurred during the same period constitute d-dimensional data Given a prediction horizon of length H, a historical observation sequence of length τ The predicted future value of the threat is The length of the lookback window is set to t≤τ and starts with the last observation y τ The end is used as the input data of the model

[0014] The complete attack data is sliced ​​and normalized on each slice. Compared with global normalization, this localized approach can better preserve the data pattern of each slice. i Split into M non-overlapping slices The mean of each slice and standard deviation The calculation is as follows:

[0015] in is the value of the slice at the tth time step. Then each slice is normalized according to its own statistical characteristics:

[0016] ε is a very small constant, used to represent the element product. Finally, the stabilized slices are reassembled according to the original time sequence of the data and used as the input data of the prediction module.

[0017] The normalized attack data As the input of the prediction module. The main goal of the prediction module is to learn a time series basis B that can explain the behavior of all time series in the group, and further use this time series basis to predict the given Corresponding Correspondingly, the time series base B can also be decomposed into the historical component B x and future component B y The work of the prediction module can be divided into three stages: learning the appropriate time series basis, calculating the correlation coefficient between the attack data and each basis, and finally realizing the prediction by weighted aggregation of the future components of the time series basis and the correlation coefficient.

[0018] First, the correlation coefficient between a set of time series and a set of time series bases is learned by using a bidirectional cross-attention block (BCAB). Given two sets of inputs and, the cross-attention block (CAB) is defined as follows:

[0019] Where MAH represents multi-head attention with H heads, and the query is Q = W q a is given, key is K=W k b is given, value is given by V = W v b is given. By exchanging information between and, BCAB is defined as follows: a (i+1) =CAB H (b (i) ,a (i) ),b (i+1) =CAB H (a (i) ,b (i) ) (a (i+1) ,b (i+1))=BCAB H (a (i) ,b (i) )

[0020] The attention mechanism allows flexible association between time series data and time series bases, ensuring that each time series can selectively focus on the most relevant time series base, and similarly, the time series base can selectively focus on the most relevant time series. Finally, the correlation coefficient between each time series base vector and each time series is calculated by x of each H head. (M) and The inner product is obtained, and the coefficient tensor is expressed as

[0021] Using the obtained correlation coefficient to make predictions, the future part Z of the time series basis y Projected into a space where linear aggregation can be performed using correlation coefficients. Through a bottleneck-like four-layer MLP Map to Then it is divided into H heads, each head has a size of N×(O / H), denoted as For each head, by calculating the The weighted sum of the correlation coefficients is used to aggregate the N time series basis vectors. The formula is defined as follows: where h∈{1,...,H} represents the head index, The size is C×H×(O / H).

[0022] Next, the H heads are connected together and passed through a four-layer fusion MLP again. Because different heads may capture different aspects of the input sequence, the fusion MLP can help converge information and improve the overall prediction performance. Finally, the mean square error (MSE) is used as the loss function to convert the stage prediction value and input data For comparison,

[0023] (x,Z x ) and (y,Z y ) is passed to the stacked BCAB blocks to obtain the correlation coefficient tensor ξ of the historical view respectively. x and the correlation coefficient tensor ξ of the future view y For each time series, x The coefficients of each time-series basis vector in are used as anchor points, ξ y The corresponding time series basis vectors in are taken as positive samples, and the remaining time series basis vectors are taken as negative samples for comparative learning. InfoNCELoss is used as the loss function to maximize ξ x and y The mutual information between them. Where δ is the temperature coefficient.

[0024] A unique statistical prediction module is introduced. Instead of using the statistical features of the original attack data to denormalize the output of the entire model, a two-layer perceptual network with an appropriate activation function is used to estimate the evolution of the future distribution of each slice. Two learnable vectors W1 and W2 are used to represent the preference weight of each variable, so that the prediction results are calculated in a weighted sum manner. The statistical prediction process can be expressed as: in Respectively represent all means and all standard deviations of the input M slices, represent the predicted mean and predicted standard deviation of future slices respectively.

[0025] While the statistics module is making predictions, the stabilized data is fed into the prediction module, which generates internal outputs. Finally, the output of the prediction module is denormalized to restore the non-stationary factors to obtain accurate prediction results. The normalization operation is performed symmetrically, the output of the prediction module is sliced, and then the slices are denormalized based on the prediction of the statistical features obtained by the statistical prediction module. The formula is defined as follows:

[0026] Finally, restore all slices in chronological order to get the final prediction result

[0027] Through the above threat prediction model, accurate prediction of long-term MCS threat prediction is achieved. The design of interpretable statistical prediction module and learnable and interpretable time series basis enhances the comprehensibility and trustworthiness of model prediction results and improves the overall security of the MCS platform.

[0028] Although the above describes the specific implementation mode of the present invention in conjunction with the drawings, it is not intended to limit the scope of protection of the present invention. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and should be included in the scope of protection of the present invention.

Claims

1. An explainable slice-level active threat prediction method in crowd sensing, characterized in that: It includes heuristic slice-level non-stationary factor processing, interpretable time series base learning, time series and time series base similarity calculation, non-stationary factor recovery and threat prediction; the specific steps are as follows: S1: For the threat data in crowd intelligence perception, slice-level normalization is performed in a heuristic way, and the statistical features of each slice are extracted to construct a statistical feature sequence. The data with non-stationary factors removed is S2: Stabilize As the standardized input of the threat prediction module, the correlation coefficient between the time series base and the time series data is calculated through the bidirectional cross attention mechanism S3: The future part Z of the time base y Project to the linear aggregation space, use the obtained correlation coefficient to weighted aggregate N time series basis vectors in dimension N, and get the smoothed output of this stage S4: Use two learnable vectors W1 and W2 to represent the preference weight of each variable, so as to calculate the future results of the statistical feature sequence in a weighted sum manner Based on the future prediction of statistical characteristics, the output is stabilized The non-stationary characteristics of the data are restored slice by slice, and the slices are aggregated in chronological order to obtain the final threat prediction results.

2. The explainable slice-level active threat prediction method in crowd intelligence perception according to claim 1 is characterized in that: The heuristic slice-level non-stationary factor processing method in S1 takes {6, 12, 24} as the slice length according to the artificially defined period (daily, weekly, etc.) in the real world, and converts the complete attack data x i Split into M non-overlapping slices The mean of each slice and standard deviation The calculation is as follows:

3. The explainable slice-level active threat prediction method in crowd intelligence perception according to claim 1 is characterized in that: The future part Z of the time base in S3 y , and the aggregation formula of the correlation coefficient is as follows: Where h∈{1,...,H} represents the head index and j is the time sequence number.

4. The explainable slice-level active threat prediction method in crowd sensing according to claim 1 is characterized in that: The module for predicting future statistical features in S4 is composed of a two-layer perception network with an appropriate activation function, and the prediction process is as follows: The formula for recovering the non-stationary characteristics of threat data is as follows:

Images