A security detection system for a server node

Abstract

This invention discloses a security detection system for server nodes, comprising a multimodal physical signal acquisition module, a clock skew measurement module, a physical field modeling and anomaly detection module, and a collaborative decision and response module. By acquiring power consumption ripple, electromagnetic radiation, acoustic signals, and thermal imaging data of nodes, and combining this with clock skew fingerprints between nodes, a node feature matrix and a dynamic adjacency matrix are constructed. A graph variational autoencoder is used to calculate the feature reconstruction error, and the rate of increase of thermal entropy for each node and the rate of increase of total cluster entropy are calculated. A three-level collaborative decision-making process is used to achieve candidate anomaly labeling, clock skew Byzantine consensus, and thermodynamic constraint verification, ultimately outputting the threat type and executing response actions. This invention eliminates the need for internal node agents, integrates multimodal physical bypass signals and inter-node measurements, and can reliably detect threats such as mining programs, firmware tampering, hardware Trojans, and clock synchronization attacks.

Description

Technical Field

[0001] This invention belongs to the field of server security detection technology, and in particular relates to a security detection system for server nodes. Background Technology

[0002] Server node security detection is a core requirement for data center and cloud computing infrastructure. Currently, mainstream server security detection technologies mainly fall into the following categories: The first is host agent-based detection schemes, which monitor processes, files, network connections, and system calls in real time by installing antivirus software, intrusion detection systems, or host monitoring tools in the server operating system. The second is network traffic-based detection schemes, which identify anomalies by analyzing the inbound and outbound traffic characteristics, communication patterns, and protocol behavior of nodes. The third is trusted computing-based security schemes, which utilize TPM / TCM chips and trusted boot metric chains to verify the integrity of firmware and the boot process. The fourth is runtime side-channel analysis schemes, such as detecting specific types of hardware trojans or mining programs through power consumption or electromagnetic signals. These technologies are widely used in data center security protection and have formed relatively mature product and standard systems.

[0003] However, existing technologies still have the following shortcomings: First, host agent-based solutions rely on the honesty of the node's operating system. Once a node is compromised, the agent software may be disabled, bypassed, or tampered with, rendering the detection results unreliable. Second, network traffic-based solutions cannot detect internal threats that do not generate abnormal traffic, such as firmware tampering, hardware Trojans, or local side-channel eavesdropping. Third, existing trusted computing solutions only cover the startup phase and lack effective monitoring of dynamic behavior during runtime. Fourth, existing single-mode side-channel detection (such as power consumption-only or electromagnetic-only) has limited coverage and cannot cope with distributed, low-intensity collaborative attacks. Fifth, existing solutions generally lack inter-node collaborative verification mechanisms, making it difficult to cross-detect attackers who tamper with sensor data or timestamps of a single node. Sixth, most existing detection systems lack hard constraint decision-making capabilities based on inviolable physical laws (such as the second law of thermodynamics), resulting in insufficient robustness against adversarial attacks. Seventh, traditional solutions often lack a complete closed loop from detection to response, especially lacking automated physical isolation and forensic support for firmware tampering and hardware Trojan threats. Therefore, there is an urgent need for a non-intrusive security detection system that does not rely on internal node proxies, can integrate multimodal physical bypass signals, and utilizes inter-node collaborative verification. Summary of the Invention

[0004] To overcome the aforementioned shortcomings of the prior art, the present invention provides a security detection system for server nodes, which solves the problems of existing technologies that rely on internal agents or single-mode signals within nodes, are easily bypassed, and lack inter-node collaboration and physical hard constraints.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A security detection system for server nodes, comprising:

[0007] The multimodal physical signal acquisition module is set up at each server node and rack location to acquire power ripple signals, electromagnetic radiation signals, acoustic signals and thermal imaging data generated during node operation, and to preprocess and extract features from the acquired signals to obtain the power consumption characteristics, electromagnetic characteristics, acoustic characteristics and thermodynamic characteristics of each node.

[0008] The clock skew measurement module utilizes the hardware timestamp function of each node's network card to measure the absolute clock skew of each node relative to the master clock and the relative clock skew between nodes through a precise time protocol, thereby establishing the clock skew fingerprint of each node and the relative skew matrix between nodes.

[0009] The physical field modeling and anomaly detection module constructs a node feature matrix based on the power consumption characteristics, electromagnetic characteristics, acoustic characteristics, thermodynamic characteristics, and absolute clock deviation, and constructs an adjacency matrix based on the thermal coupling strength, electromagnetic coupling strength, and clock coupling strength between nodes. The node feature matrix and adjacency matrix are input into the trained graph variational autoencoder to calculate the feature reconstruction error of each node, and to calculate the thermal entropy growth rate of each node and the total entropy growth rate of the cluster.

[0010] The collaborative decision and response module reconstructs the error marker candidate abnormal nodes based on the features, performs inter-node physical field consistency verification and thermodynamic constraint verification on the candidate abnormal nodes, outputs the threat type and confidence level, and executes response actions based on the decision results.

[0011] Preferably, the multimodal physical signal acquisition module includes:

[0012] A Hall current sensor is installed at the output of the CPU power supply VRM to collect the CPU dynamic current ripple.

[0013] A near-field electromagnetic probe is placed near the memory slot to collect electromagnetic radiation signals.

[0014] A MEMS ultrasonic microphone mounted on the fan frame is used to collect the infrasound and ultrasonic components in the fan noise.

[0015] Infrared thermal imaging arrays are installed at the air inlet on the front panel and the air outlet at the rear of the server to collect two-dimensional thermal distribution data.

[0016] The preprocessing and feature extraction include: low-pass filtering and wavelet packet decomposition of power consumption ripple, wavelet denoising and spectral peak detection of electromagnetic radiation, spectral subtraction and empirical mode decomposition of acoustic signals, and median filtering and temperature gradient entropy calculation of thermal imaging data.

[0017] Preferably, in the clock skew measurement module: each node sends PTP messages to all other nodes at a preset period and records the hardware timestamp, and calculates the relative clock skew; under the safe state at the initial stage of cluster deployment, the absolute clock skew sequence of a preset duration is continuously collected, the long-term drift trend is extracted using an exponentially weighted moving average, and a residual distribution is established using a Gaussian mixture model as the clock skew fingerprint; the singular value decomposition principal components of the relative skew matrix of all nodes are calculated as the clock consistency fingerprint; and a preset tolerance is used to determine clock skew anomalies.

[0018] Preferably, the thermal coupling strength is calculated based on the mutual information entropy between the outlet temperature of a node and the inlet temperature of an adjacent node; the electromagnetic coupling strength is calculated based on the coherence coefficient of the electromagnetic spectrum between nodes; the clock coupling strength is calculated based on the exponential decay function of the relative clock deviation and the historical mean; the elements of the adjacency matrix are the maximum values ​​of the thermal coupling strength, electromagnetic coupling strength, and clock coupling strength, and the corresponding element is set to zero when the maximum value is lower than a preset threshold.

[0019] Preferably, the graph variational autoencoder includes an encoder, an inner product decoder, and a graph generator decoder; the encoder is a multi-layer graph convolutional network used to map the node feature matrix and adjacency matrix to the mean and variance of latent variables; the inner product decoder is used to reconstruct the adjacency matrix; the graph generator decoder is used to reconstruct the node feature matrix; the data used to train the graph variational autoencoder is data continuously collected by the cluster for a preset duration under a confirmed no-threat state, and the loss function during training includes structural reconstruction loss, feature reconstruction loss, and KL divergence regularization term.

[0020] Preferably, the rate of increase of thermal entropy is calculated according to the following formula:

[0021] ,in, For the specific heat of air, The fan mass flow rate is estimated based on the fan's operating parameters. , The temperatures at the node's air inlet and outlet are denoted as Δt, and Δt is the time interval. The total entropy increment of the cluster is the sum of the growth rates of the thermal entropy of all nodes. An autoregressive integral moving average model is used to establish the normal prediction interval for the growth rate of the total entropy of the cluster.

[0022] Preferably, the collaborative decision and response module performs a three-level collaborative decision:

[0023] Level 1: Input the real-time node features into the graph variational autoencoder, calculate the feature reconstruction error, and if the feature reconstruction error exceeds the threshold dynamically determined by the exponential weighted moving average, then mark the node as a candidate abnormal node.

[0024] Level 2: For candidate abnormal nodes, extract the edge feature differences between them and their neighboring nodes, and perform Byzantine consensus on clock deviation between nodes: If the difference between the relative clock deviation independently measured by more than a preset proportion of neighboring nodes and the absolute clock deviation conversion value reported by the node itself is greater than a preset tolerance, it is determined that there is timestamp tampering or clock synchronization attack, calculate the consistency score, and mark the node as a highly suspicious node when the consistency score is lower than a preset threshold.

[0025] Level 3: For highly suspicious nodes, compare the changing trends of their thermal entropy growth rate with the total entropy growth rate of the cluster, and perform time-series causal relationship analysis on the thermal entropy growth rate and power consumption to determine whether there is sensor tampering or hidden computing tasks.

[0026] Preferably, the collaborative decision and response module further includes a random forest classifier. The inputs of the random forest classifier are the feature reconstruction error calculated at the first level, the consistency score calculated at the second level, and the thermodynamic deviation calculated at the third level. The outputs a threat type label and a confidence level. The threat type label includes: normal, mining program, ransomware, DDoS botnet, firmware tampering, hardware Trojan, side-channel eavesdropping, and clock synchronization attack.

[0027] Preferably, the response actions include: recording alarms, notifying the administrator, remotely restarting the node via BMC, physically powering off and isolating the node, and triggering memory dump for evidence collection; when the threat type is firmware tampering or hardware Trojan, the collaborative decision and response module automatically performs physical power-off and notifies the replacement of hardware; the collaborative decision and response module also retains the original sensor data and clock deviation matrix for a preset time before and after the alarm time as digital forensic evidence.

[0028] Preferably, in the physical field modeling and anomaly detection module, the time window length used to construct the node feature matrix is ​​a first preset duration, and the sliding step size is a second preset duration; the multimodal physical signal acquisition module, clock deviation measurement module, physical field modeling and anomaly detection module, and collaborative decision and response module exchange data through an out-of-band management network, and all sensor data are timestamped by the node's BMC during acquisition; when the detection result is confirmed as a false alarm, the data within the false alarm time window is used as a normal sample, and the model parameters of the graph variational autoencoder are incrementally updated.

[0029] The technical effects and advantages of the security detection system for server nodes of the present invention are as follows:

[0030] 1. This invention directly collects power consumption ripple signals, electromagnetic radiation signals, acoustic signals, and thermal imaging data generated by server nodes during operation through multimodal physical signal acquisition modules deployed at nodes and rack locations, without requiring any proxy software to be installed in the node's operating system. All sensors are independent of the node's software stack, making the detection system imperceptible to attackers, and preventing them from bypassing detection by tampering with the operating system or applications. Compared to traditional security detection schemes that rely on host proxies or network traffic, this invention fundamentally eliminates the risk of detection failure after the node itself is compromised.

[0031] 2. The clock skew measurement module in this invention utilizes the hardware timestamp function of each node's network interface card (NIC) to measure the absolute clock skew of each node relative to the master clock, as well as the relative clock skew between nodes, through a precise time protocol. It then establishes a clock skew fingerprint for each node and a relative skew matrix between nodes. The measurement results between nodes are independent of any single node's report, and the Byzantine consensus mechanism effectively identifies timestamp tampering or clock synchronization attacks. This feature ensures that even if an attacker completely controls a node and falsifies its clock skew report, they cannot simultaneously deceive the independent measurement results of all neighboring nodes, thus providing an immutable hardware-level trust foundation.

[0032] 3. This invention constructs a node feature matrix by combining the power consumption characteristics, electromagnetic characteristics, acoustic characteristics, thermodynamic characteristics, and absolute clock skew of each node. Simultaneously, it dynamically constructs an adjacency matrix based on the thermal coupling strength, electromagnetic coupling strength, and clock coupling strength between nodes, and uses a graph variational autoencoder to uniformly model the cluster's physical field. This model not only captures the internal state of each node but also learns the normal physical coupling relationships between nodes. When an attacker distributes malicious payloads across multiple nodes to evade single-node detection, the thermal coupling, electromagnetic coherence, or clock skew correlation between nodes will inevitably deviate from the normal pattern, thus being captured by the edge feature reconstruction error of the graph variational autoencoder. Compared to schemes that only monitor a single physical quantity or only detect a single node, this invention significantly improves the detection sensitivity and robustness against distributed cooperative attacks.

[0033] 4. This invention calculates the rate of increase of thermal entropy for each node and the rate of increase of total cluster entropy, and uses an autoregressive integral moving average model to establish a normal prediction interval. Any computationally intensive malicious load (such as mining programs, ransomware, or DDoS attacks) will inevitably lead to an increase in the rate of increase of total system entropy. This law is determined by the second law of thermodynamics and cannot be changed by any software-level countermeasures. Even if attackers tamper with some sensor data or use extremely low-intensity covert loads, thermodynamic constraints still provide the final physical basis for judgment. Compared with existing methods that rely solely on statistical anomalies, this invention has stronger anti-spoofing capabilities.

[0034] 5. This invention employs a three-level collaborative decision-making process: the first level uses feature reconstruction errors based on a graph variational autoencoder to label candidate abnormal nodes; the second level calculates consistency scores through Byzantine consensus of clock skew between nodes and edge feature differences to identify timestamp tampering; the third level utilizes thermodynamic constraints to verify the existence of sensor spoofing or covert computation tasks. These three levels of decision-making complement and corroborate each other, effectively avoiding the limitations of a single detection dimension. A random forest classifier integrates the outputs of the three levels, outputting specific threat type labels and confidence levels, providing operations and maintenance personnel with clear and actionable alarm information.

[0035] 6. When a detection result is confirmed as a false alarm, this invention can treat the data within the false alarm time window as normal samples and incrementally update the model parameters of the graph variational autoencoder. This mechanism enables the detection system to adapt to long-term dynamic changes such as server node aging, environmental changes, and business model migration, avoiding the problem of performance degradation of a fixed model over time.

[0036] 7. This invention supports multiple response actions, including recording alarms, notifying administrators, remotely restarting nodes via BMC, physically powering off and isolating nodes, and triggering memory dump for evidence collection. It also retains original sensor data and clock deviation matrices for a preset duration before and after the alarm as digital forensic evidence. For high-confidence threats such as firmware tampering or hardware Trojans, the system can automatically perform a physical power-off and notify for hardware replacement, achieving a complete security closed loop from detection to response. Attached Figure Description

[0037] Figure 1 This is a flowchart of a security detection system for server nodes proposed in this invention;

[0038] Figure 2 This is a multimodal physical signal acquisition and feature extraction diagram of a server node security detection system proposed in this invention;

[0039] Figure 3 This invention relates to a clock skew measurement and Byzantine consensus graph for a server node security detection system. Detailed Implementation

[0040] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0041] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include," "contain," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "includes..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0042] refer to Figure 1-3 This invention discloses a security detection system for server nodes. The system includes: a multimodal physical signal acquisition module, installed at each server node and rack location, used to acquire power ripple signals, electromagnetic radiation signals, acoustic signals, and thermal imaging data generated during node operation, and extract the power consumption characteristics, electromagnetic characteristics, acoustic characteristics, and thermodynamic characteristics of each node; a clock skew measurement module, utilizing the hardware timestamp function of each node's network card, to measure the absolute clock skew of each node relative to the master clock and the relative clock skew between nodes through a precise time protocol, establishing a clock skew fingerprint for each node and a relative skew matrix between nodes; and physical field modeling and anomaly detection. The detection module constructs a node feature matrix based on power consumption characteristics, electromagnetic characteristics, acoustic characteristics, thermodynamic characteristics, and absolute clock deviation. It also constructs an adjacency matrix based on the thermal coupling strength, electromagnetic coupling strength, and clock coupling strength between nodes. The node feature matrix and adjacency matrix are input into a trained graph variational autoencoder to calculate the feature reconstruction error of each node, and to calculate the rate of increase of thermal entropy for each node and the rate of increase of total cluster entropy. The collaborative decision and response module marks candidate abnormal nodes based on the feature reconstruction error, performs physical field consistency verification and thermodynamic constraint verification on the candidate abnormal nodes, outputs the threat type and confidence level, and executes response actions based on the decision results. This invention does not rely on internal node proxies. Through multimodal physical bypass signal fusion, mutual measurement of clock deviation between nodes, graph neural network physical field modeling, and thermodynamic constraint decision-making, it achieves highly reliable, bypass-resistant, and non-intrusive security detection of various threats in server clusters, such as mining programs, firmware tampering, hardware trojans, and clock synchronization attacks.

[0043] Example 1

[0044] Purpose of implementation: To verify the detection capability of this invention for single-node covert mining programs (using only a small number of CPU cores and running late at night), especially the sensitivity of multimodal physical signals to lightweight computing loads.

[0045] System Implementation: The aforementioned 8-node cluster was used, with the XMRig mining software running on node A, utilizing only 4 CPU cores. The system was scheduled to run between 2:00 AM and 4:00 AM to simulate an attacker attempting to evade traditional CPU utilization-based detection. The rest of the system remained operational, including the multimodal physical signal acquisition module, clock skew measurement module, physical field modeling and anomaly detection module, and collaborative decision-making and response module, all of which were enabled.

[0046] Implementation process:

[0047] Real-time acquisition of power consumption ripple at node A: a new frequency band energy appears near 1.2kHz (wavelet packet band 8 energy increases by 40%); electromagnetic radiation shows a new peak at 1.5GHz; the instantaneous frequency fluctuation of the ultrasonic component (40kHz) increases; thermal imaging shows that the local temperature gradient entropy in the CPU area increases from 0.65 to 0.89.

[0048] The absolute clock deviation of node A accumulated a drift of +23 ns (relative to the GPS disciplined clock) within 15 minutes after mining began, while the drifts of adjacent nodes B and C were both less than ±5 ns. The relative deviations measured by nodes B and C from node A were +24 ns and +21 ns, respectively, and the deviation change vectors deviated significantly from the historical pattern.

[0049] The graph variational autoencoder in the physics field modeling and anomaly detection module calculates the feature reconstruction error of node A as 0.87. The dynamic threshold (mean of the past 24 hours 0.23 + 3 × standard deviation 0.05 = 0.38) exceeds the threshold, so node A is marked as a candidate anomaly node.

[0050] The collaborative decision and response module performs the second-level consistency check: the edge feature differences between node A and its neighbors B and C are 0.25 and 0.31 respectively, and the consistency score C = 0.69 (higher than 0.6), and the Byzantine consensus is not triggered (because the node itself has not tampered with the clock).

[0051] Level 3 thermodynamic verification: The entropy growth rate of node A increased from 0.12 W / K·s to 0.31 W / K·s, while the overall cluster entropy growth rate only increased from 2.1 to 2.3, which did not exceed the upper limit of the ARIMA prediction range of 2.4. Granger causality test showed that the change in entropy growth rate lagged behind the change in power consumption (p=0.72), consistent with normal load characteristics.

[0052] The random forest classifier (100 decision trees, maximum depth 10) in the collaborative decision and response module receives a reconstruction error of 0.87, a consistency score of 0.69, a thermodynamic deviation of 0.2, and outputs a threat type label "mining program" with a confidence level of 0.94.

[0053] Implementation Results: The total delay from the start of the mining program to the system outputting an alarm was approximately 90 seconds. The system sent a command to the BMC via the Redfish interface to isolate and restart node A, while retaining all original sensor data and clock skew matrix from the preceding and following 5 minutes as digital forensic evidence. This embodiment demonstrates that even when the mining program uses only 4 cores and runs during low-load periods at night, this invention can still achieve high-confidence detection through multimodal fusion features of power ripple, electromagnetic radiation, and thermal imaging, without relying on any internal agent within the node.

[0054] Example 2

[0055] Purpose of implementation: To verify the ability of this invention to detect UEFI firmware tampering in supply chain attacks, especially when tampering causes abnormal node power consumption and thermal characteristics but does not change the visible behavior of the operating system, whether the system can detect the anomaly through consensus on clock deviation between nodes and thermodynamic constraints.

[0056] System Implementation: Malicious code (modifying the ACPI table to maintain high CPU voltage even under low load) is implanted in the UEFI firmware of Node B, simulating a firmware backdoor implanted by an attacker during manufacturing or shipping. The rest of the system is the same as in Example 1, and all modules function normally.

[0057] Implementation process:

[0058] The power ripple characteristics show that even when node B is in an idle state, the DC component of its power ripple is 15% higher than normal, and the switching frequency harmonic amplitude is abnormal.

[0059] Electromagnetic radiation: Abnormal modulation sidebands appear at 300MHz, which do not match the spectrum pattern during normal firmware startup.

[0060] Thermal imaging: Inlet temperature 25°C, outlet temperature 38°C (normally should be 32°C), thermal relaxation time constant shortened from 120 seconds to 85 seconds.

[0061] Clock skew: The long-term drift slope of the absolute clock skew at node B changed from +8ns per day to +23ns per day, which is seriously inconsistent with the crystal oscillator aging model.

[0062] The feature reconstruction error of node B in the graph variational autoencoder is 0.95, which exceeds the dynamic threshold of 0.32, and is therefore marked as a candidate anomaly.

[0063] Level 2: The relative clock deviations independently measured by Node B's neighboring nodes C, D, and E differ from Node B's self-reported value by up to 78 ns (preset tolerance 50 ns). Since more than half (3 / 4) of the neighbors' measurements exceed this limit, the Byzantine consensus determines that Node B has undergone timestamp tampering or a clock synchronization attack. A consensus score of C = 0.28 (below 0.6) is calculated, and Node B is marked as a highly suspicious node.

[0064] Level 3: The thermal entropy growth rate of node B (0.18W / K·s) and the total entropy growth rate of the cluster (2.2W / K·s) show inconsistent trends (node ​​B's entropy increases but the total entropy increase is not synchronized), which is determined to be sensor tampering or firmware-level attack.

[0065] The random forest classifier receives a reconstruction error of 0.95, a consistency score of 0.28, and a thermodynamic deviation of 0.7. It outputs a threat type label of "firmware tampering" with a confidence level of 0.98.

[0066] Implementation Results: The system automatically performs a physical power outage to isolate node B and notifies the administrator to replace the motherboard via the out-of-band management network. This embodiment demonstrates that firmware backdoors that traditional antivirus software cannot detect can be reliably detected through the inter-node clock skew consensus (utilizing the characteristic that hardware timestamps cannot be tampered with by software) and thermodynamic constraints (the rate of entropy increase does not match the overall cluster) in this invention.

[0067] Example 3

[0068] Purpose of implementation: To verify the ability of this invention to detect hardware Trojans implanted inside the CPU. These Trojans are activated only when receiving specific network packets and their presence is exposed through the narrowband characteristics of electromagnetic radiation.

[0069] System Implementation: A hardware Trojan is simulated inside the CPU of node C (via an FPGA prototyping platform). This Trojan is activated upon receiving a specific UDP packet and steals the encryption key by modifying the L3 cache data. The rest of the system remains unchanged.

[0070] Implementation process:

[0071] During normal operation, the electromagnetic radiation spectrum of node C is stable. When the attacker sends a trigger packet, the Trojan activity causes a narrowband peak in the 1.8GHz band, which is 12dB higher than the background amplitude. This band corresponds to the CPU's internal ring bus.

[0072] The power ripple exhibits irregular pulses (pulse width approximately 200ns), which is significantly different from the normal memory access mode.

[0073] The infrasound component (12Hz) in the acoustic signal exhibits periodic fluctuations, corresponding to the periodic switching of the Trojan state machine.

[0074] The reconstruction error of the graph variational autoencoder computation node C is 0.76, and the dynamic threshold is 0.41. Nodes exceeding the threshold are marked as candidate anomalies.

[0075] Level 2: No anomalies were detected in neighboring nodes (the Trojan only affected the internal workings of node C), node C itself did not tamper with the clock, and the consistency score C = 0.88 (higher than 0.6).

[0076] Level 3: The rate of increase in thermal entropy at node C was only 0.03 W / K·s, which did not exceed the ARIMA prediction range, and the Granger causality test showed no significant causal relationship.

[0077] The random forest classifier received a reconstruction error of 0.76, a consistency score of 0.88, and a thermodynamic deviation of 0.1. It outputs a threat type label of "hardware Trojan" with a confidence level of 0.85.

[0078] Implementation Results: The system logs alarms and recommends that maintenance personnel use JTAG debugging for in-depth analysis. This example demonstrates that even without significant thermodynamic and clock deviations, the narrowband characteristics of electromagnetic radiation can still be captured by the multimodal physical signal acquisition module, and hardware Trojans can be detected through the reconstruction error of the graph variational autoencoder, verifying the unique value of electromagnetic features in hardware-level threat detection.

[0079] Example 4

[0080] Purpose of implementation: To verify the detection capability of this invention against PTP clock synchronization attacks, especially when an attacker compromises a node and forges its clock deviation report, whether the system can independently detect the deception through mutual measurements between nodes.

[0081] System Implementation: An attacker compromised node D through a vulnerability, modifying its PTP protocol stack to cause node D to broadcast false clock synchronization messages to other nodes (claiming that its clock deviation from the master clock is less than 1 ns), attempting to disrupt the time base of the entire cluster. All system modules function normally.

[0082] Implementation process:

[0083] Node D's own clock skew measurement module reported an error value (absolute skew <1ns) due to an attack.

[0084] The relative clock offsets of other nodes (E, F, G, H) to node D were measured independently: node E measured +340ns, node F measured +355ns, node G measured -28ns (abnormal, possibly due to measurement error), and node H measured +348ns. Most (3 / 4) of the measurements were approximately +350ns.

[0085] After converting the absolute deviation reported by node D into a relative deviation (assuming the master clock is stable), the difference between it and the neighbor's measurement exceeds a preset tolerance of 50 ns. Since the neighbor's measurement exceeds a preset proportion (3 / 4 > preset proportion 0.5), the Byzantine consensus determines that node D has had its timestamp tampered.

[0086] The power consumption, electromagnetic, acoustic, and thermal characteristics of node D are all normal (the attack did not change the computational load), so the first-level reconstruction error is only 0.12, which does not exceed the dynamic threshold of 0.30.

[0087] However, with a second-level consistency score of C=0.20 (below 0.6), the node is directly marked as a highly suspicious node without the need for third-level verification.

[0088] The random forest classifier received a reconstruction error of 0.12, a consistency score of 0.20, and a thermodynamic deviation of 0 (level 3 was not triggered). It outputs a threat type label "clock synchronization attack" with a confidence level of 0.99.

[0089] Implementation Results: The system isolates node D from the PTP clock domain and automatically switches to a backup clock source, while simultaneously notifying the administrator via the out-of-band management network. This embodiment demonstrates that the clock skew measurement module can independently detect clock synchronization attacks by utilizing hardware timestamp information measured between nodes, without relying on node reports themselves, showcasing the unique advantages of the "mutual supervision between nodes" mechanism of this invention.

[0090] Example 5

[0091] Purpose of implementation: To verify the detection capability of this invention against distributed collaborative attacks, i.e., when attackers distribute the mining load to multiple nodes, each node uses only a small amount of resources, making the characteristics of a single node inconspicuous, but the physical field coupling relationship between nodes (thermal coupling, electromagnetic coherence, clock deviation correlation) will still expose anomalies.

[0092] System Implementation: Lightweight mining programs are installed on nodes E, F, and G respectively. Each node uses only one CPU core, and the mining execution sequence of the three nodes is staggered (to make the load fluctuation of a single node discontinuous). All modules of the system work normally.

[0093] Implementation process:

[0094] Feature reconstruction error for a single node: 0.31 for node E (slightly exceeds the dynamic threshold of 0.30), 0.29 for node F (not exceeded), and 0.32 for node G (exceeds the threshold). Relying solely on single-node detection may result in misclassification as noise.

[0095] However, the graph variational autoencoder simultaneously reconstructs the edge features between nodes: the thermal coupling strength (mutual information entropy) between nodes EF, EG, and FG increases from 0.15~0.20 under normal conditions to 0.35~0.42, because the outlet temperature of the three nodes increases simultaneously and the heat recirculation is enhanced.

[0096] Electromagnetic coupling strength: The three nodes show a common peak in the 2.4GHz band (the mining algorithm uses the same instruction pattern).

[0097] Clock skew: The absolute skew of nodes E, F, and G all show positive drift (+12ns, +15ns, +13ns) within the same time window, while the drift of other nodes is less than 3ns.

[0098] Second-level consistency check: Although the clock deviation reported by each node is consistent with the neighbor's measurement (no tampering), the singular principal components of the relative deviation matrix of the three nodes deviate from the benchmark (the variance contribution rate of the first principal component drops from 85% to 62%), and the system detects "group clock anomaly".

[0099] Level 3: The rate of increase of the total entropy of the cluster increased from 2.1 W / K·s to 2.7 W / K·s, exceeding the upper limit of the ARIMA prediction range of 2.5. Furthermore, the Granger causality test showed that the change in the rate of increase of entropy led the change in the power consumption of any single node (p=0.04), indicating the existence of an unauthorized overall computing load.

[0100] The random forest classifier integrates population features (using the reconstruction error, consistency score, and thermodynamic deviation of the three nodes as joint inputs) to output the threat type label "mining program (multi-node collaboration)" with a confidence level of 0.96.

[0101] Implementation Results: The system simultaneously isolates nodes E, F, and G, while retaining all relevant sensor data as evidence. This embodiment verifies the ability of the physical field coupling modeling (thermal coupling strength, electromagnetic coupling strength, and clock coupling strength) between nodes and the reconstruction capability of the graph variational autoencoder for edge features in this invention. It can effectively detect distributed, low-intensity cooperative attacks, which traditional single-node detection methods cannot identify.

[0102] Comparative Example 1

[0103] Purpose of implementation: To demonstrate, in contrast to Example 1 (single-node mining), the limitations of traditional detection methods that rely on operating system views when facing adversarial tactics such as process hiding and traffic encryption.

[0104] System Implementation: A traditional security detection solution is deployed on the same 8-node cluster: each server is equipped with ClamAV antivirus software (with daily virus database updates) and Osquery host monitoring tool (monitoring processes, files, and network connections), and configured with CPU usage-based threshold alarms (alarms are triggered when CPU usage exceeds 20%). This solution does not include any of the physical signal acquisition modules, clock deviation measurement modules, physical field modeling and anomaly detection modules, or collaborative decision-making and response modules of this invention.

[0105] Implementation Process: Similar to Example 1, the XMRig mining software was run on node A, utilizing only 4 CPU cores, from 2:00 AM to 4:00 AM. The mining program employed process hiding technology (hookreaddir system call) to prevent Osquery from listing suspicious processes. Simultaneously, the mining program encapsulated mining pool communication within HTTPS traffic, did not use any known malicious domains, and neither ClamAV nor traditional IDS generated alerts. CPU utilization only increased by approximately 12%, below the 20% threshold, therefore no alerts were triggered.

[0106] Implementation Results: The mining program ran continuously for 7 days, consuming approximately 150 kWh of power, without being detected. The anomaly was only discovered during manual inspection by maintenance personnel. This comparison demonstrates that traditional detection methods based on host proxies and network traffic are easily bypassed by anti-piracy measures such as process hiding and traffic encryption. In contrast, this invention utilizes physical bypass signals (power consumption, electromagnetic, acoustic, and thermal radiation) completely independent of the operating system view; even if the mining program hides its processes and traffic, its physical characteristics cannot be concealed.

[0107] Comparative Example 2

[0108] Purpose of implementation: In contrast to Example 5 (multi-node collaborative mining), this invention demonstrates the inadequacy of detection methods that rely solely on a single physical modality (such as power consumption) when dealing with distributed, low-intensity attacks, thereby verifying the necessity of multimodal fusion and inter-node coupling analysis in this invention.

[0109] System Implementation: A simplified detection system is deployed on the same 8-node cluster. It uses only the Hall current sensor on each node to collect power consumption ripple and employs a simple threshold detection algorithm (triggers an alarm when a node's power consumption exceeds 20% of the historical average). This system does not include electromagnetic probes, microphones, thermal imaging arrays, clock skew measurement modules, graph variational autoencoders, or inter-node coupling analysis capabilities.

[0110] Implementation Process: Similar to Example 5, lightweight mining programs were run on nodes E, F, and G, with each node using only one CPU core and staggered execution order. The power consumption increase of a single node was only 5-8%, far below the 20% alarm threshold. The frequency domain characteristics of the power ripple highly overlapped with normal fluctuations (such as system background tasks and disk I / O), making them indistinguishable. No alarms were triggered by the system.

[0111] Implementation Results: The three nodes ran the mining program continuously for over a week without being detected. This comparative example demonstrates that a single power consumption mode cannot detect distributed, low-intensity coordinated attacks because the signal variations of each node are submerged in normal fluctuations. However, this invention, by fusing thermal imaging (capturing thermal coupling), electromagnetic spectrum analysis (capturing coherent peaks), clock skew (capturing group drift), and graph variational autoencoder reconstruction of inter-node edge features, can detect anomalies from changes in the overall physical field of the cluster, significantly improving detection sensitivity and robustness.

[0112] The detection results of the five embodiments are as follows: Embodiment 1 (mining program) detection delay approximately 90 seconds, confidence level 0.94; Embodiment 2 (firmware tampering) detection delay approximately 60 seconds, confidence level 0.98; Embodiment 3 (hardware Trojan) detection delay approximately 120 seconds, confidence level 0.85; Embodiment 4 (clock synchronization attack) detection delay approximately 30 seconds, confidence level 0.99; Embodiment 5 (multi-node mining) detection delay approximately 150 seconds, confidence level 0.96. No false positives or missed detections were observed in any of the embodiments. Comparative Examples 1 and 2 failed to detect the corresponding threats, verifying the significant advancement of this invention compared to the prior art.

[0113] Compared with Examples 1-5 and Comparative Example 1, Examples 1-5 of this invention were verified against five typical threat scenarios: mining programs, firmware tampering, hardware Trojans, clock synchronization attacks, and multi-node collaborative mining. In Example 1, the system detected covert mining within 90 seconds with a confidence level of 0.94 by collecting the power consumption ripple (40% increase in energy in the 1.2kHz band), electromagnetic radiation (a new peak at 1.5GHz), ultrasonic component fluctuations, and the increase in thermal imaging temperature gradient entropy (0.65→0.89) of node A, combined with the feature reconstruction error calculated by the graph variational autoencoder (0.87 exceeding the threshold of 0.38). Example 2 addresses UEFI firmware tampering. The system not only detected a 15% increase in the DC component of power consumption and an abnormal 300MHz sideband in electromagnetic radiation, but more importantly, the clock deviation slope of node B jumped from +8ns / day to +23ns / day. Furthermore, the relative deviation independently measured by neighboring nodes differed from the node's self-reported value by 78ns (exceeding the 50ns tolerance). The Byzantine consensus gave a consistency score of only 0.28, ultimately determining firmware tampering with a confidence level of 0.98. In Example 3, the hardware Trojan only activated upon receiving specific network packets, causing a 12dB narrowband peak in 1.8GHz electromagnetic radiation and 200ns irregular pulses in power consumption ripple. Although there were no obvious thermodynamic or clock anomalies, the reconstruction error of 0.76 still exceeded the threshold, resulting in a hardware Trojan alarm output with a confidence level of 0.85. In Example 4, the attacker tampered with the PTP protocol stack of node D, causing it to falsely report clock skew. However, the relative deviations measured independently by nodes E to H (approximately +350ns) were significantly inconsistent with the self-reported value of node D (<1ns), exceeding the preset proportion of neighbor measurements. The consistency score was as low as 0.20, and the clock synchronization attack was detected directly with a confidence level of 0.99 without thermodynamic verification, with a delay of only 30 seconds. In Example 5, the attacker distributed the mining load to nodes E, F, and G, with each node using only one core and staggered timing. The reconstruction error of a single node only slightly exceeded the threshold, but the thermal coupling strength between nodes increased from 0.15~0.20 to 0.35~0.42. The electromagnetic coherence peaks appeared together in the 2.4GHz band, the principal component of the relative deviation matrix deviated from the baseline (variance contribution rate 85%→62%), and the growth rate of the total entropy of the cluster (2.7W / K·s) exceeded the upper limit of the ARIMA prediction interval of 2.5. Finally, multi-node collaborative mining was detected with a confidence level of 0.96. All five implementations achieved successful detection with no false negatives or missed positives, and the detection delay was between 30 and 150 seconds.

[0114] Comparative Example 1 uses a traditional host-agent-based security detection scheme (ClamAV antivirus software, Osquery host monitoring tool, and CPU usage threshold alarms), which completely fails in the same mining scenario (Node A in Example 1). The mining program hides its process by hooking the readdir system call, and Osquery cannot list suspicious processes; the mining pool communication is encapsulated in HTTPS traffic and does not use known malicious domain names, so neither ClamAV nor traditional IDS issues alarms; the CPU usage only increases by about 12%, which is below the 20% threshold, and the system does not respond. This mining program runs continuously for 7 days, consuming about 150kWh of electricity, and remains undetected. The comparison shows that traditional solutions rely on operating system views and network characteristics, making them easily bypassed by anti-interference measures such as process hiding and traffic encryption; while this invention utilizes physical bypass information such as unavoidable power consumption ripple, electromagnetic radiation, acoustic signals, thermal radiation, and inter-node clock deviations that are unavoidable during server node operation, completely independent of the node's internal software stack. Attackers cannot simultaneously forge multimodal physical characteristics and inter-node clock coupling relationships. Especially for threats that traditional solutions struggle to cover, such as firmware tampering, hardware trojans, and clock synchronization attacks, this invention achieves highly reliable, non-intrusive security detection through unified physical field modeling, inter-node measurement, and thermodynamic constraint decision-making.

[0115] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of protection of the claims.

[0116] In conclusion, the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A security detection system for server nodes, characterized in that, include: The multimodal physical signal acquisition module is set up at each server node and rack location to acquire power ripple signals, electromagnetic radiation signals, acoustic signals and thermal imaging data generated during node operation, and to preprocess and extract features from the acquired signals to obtain the power consumption characteristics, electromagnetic characteristics, acoustic characteristics and thermodynamic characteristics of each node. The clock skew measurement module utilizes the hardware timestamp function of each node's network card to measure the absolute clock skew of each node relative to the master clock and the relative clock skew between nodes through a precise time protocol, thereby establishing the clock skew fingerprint of each node and the relative skew matrix between nodes. The physical field modeling and anomaly detection module constructs a node feature matrix based on the power consumption characteristics, electromagnetic characteristics, acoustic characteristics, thermodynamic characteristics, and absolute clock deviation, and constructs an adjacency matrix based on the thermal coupling strength, electromagnetic coupling strength, and clock coupling strength between nodes. The node feature matrix and adjacency matrix are input into the trained graph variational autoencoder to calculate the feature reconstruction error of each node, and to calculate the thermal entropy growth rate of each node and the total entropy growth rate of the cluster. The collaborative decision and response module reconstructs the error marker candidate abnormal nodes based on the features, performs inter-node physical field consistency verification and thermodynamic constraint verification on the candidate abnormal nodes, outputs the threat type and confidence level, and executes response actions based on the decision results.

2. The security detection system for a server node as described in claim 1, characterized in that, The multimodal physical signal acquisition module includes: A Hall current sensor is installed at the output of the CPU power supply VRM to collect the CPU dynamic current ripple. A near-field electromagnetic probe is placed near the memory slot to collect electromagnetic radiation signals. A MEMS ultrasonic microphone mounted on the fan frame is used to collect the infrasound and ultrasonic components in the fan noise. Infrared thermal imaging arrays are installed at the air inlet on the front panel and the air outlet at the rear of the server to collect two-dimensional thermal distribution data. The preprocessing and feature extraction include: low-pass filtering and wavelet packet decomposition of power consumption ripple, wavelet denoising and spectral peak detection of electromagnetic radiation, spectral subtraction and empirical mode decomposition of acoustic signals, and median filtering and temperature gradient entropy calculation of thermal imaging data.

3. The security detection system for a server node as described in claim 1, characterized in that, In the clock skew measurement module: each node sends PTP messages to all other nodes according to a preset period and records the hardware timestamp, and calculates the relative clock skew; in the safe state of the initial deployment of the cluster, the absolute clock skew sequence of a preset duration is continuously collected, the long-term drift trend is extracted by exponential weighted moving average, and the residual distribution is established as the clock skew fingerprint by Gaussian mixture model. Calculate the singular value decomposition principal components of the relative deviation matrix for all nodes, and use them as clock consistency fingerprints; Clock deviation anomalies are determined using a preset tolerance.

4. The security detection system for a server node as described in claim 1, characterized in that, The thermal coupling strength is calculated based on the mutual information entropy between the outlet temperature of a node and the inlet temperature of a neighboring node; the electromagnetic coupling strength is calculated based on the coherence coefficient of the electromagnetic spectrum between nodes; the clock coupling strength is calculated based on the exponential decay function of the relative clock deviation and the historical mean; the elements of the adjacency matrix are the maximum values ​​of the thermal coupling strength, electromagnetic coupling strength, and clock coupling strength, and the corresponding element is set to zero when the maximum value is lower than a preset threshold.

5. The security detection system for a server node as described in claim 1, characterized in that, The graph variational autoencoder includes an encoder, an inner product decoder, and a graph generator decoder; the encoder is a multi-layer graph convolutional network used to map the node feature matrix and adjacency matrix to the mean and variance of latent variables; the inner product decoder is used to reconstruct the adjacency matrix; and the graph generator decoder is used to reconstruct the node feature matrix. The data used to train the graph variational autoencoder is data collected continuously for a preset duration by the cluster under a confirmed no-threat state. The loss function during training includes structural reconstruction loss, feature reconstruction loss, and KL divergence regularization term.

6. The security detection system for a server node as described in claim 1, characterized in that, The rate of increase of thermal entropy is calculated according to the following formula: ,in, For the specific heat of air, The fan mass flow rate is estimated based on the fan's operating parameters. , The temperatures at the node's air inlet and outlet are denoted as Δt, and Δt is the time interval. The total entropy increment of the cluster is the sum of the growth rates of the thermal entropy of all nodes. An autoregressive integral moving average model is used to establish the normal prediction interval for the growth rate of the total entropy of the cluster.

7. The security detection system for a server node as described in claim 1, characterized in that, The collaborative decision-making and response module performs a three-level collaborative decision-making process: Level 1: Input the real-time node features into the graph variational autoencoder, calculate the feature reconstruction error, and if the feature reconstruction error exceeds the threshold dynamically determined by the exponential weighted moving average, then mark the node as a candidate abnormal node. Level 2: For candidate abnormal nodes, extract the edge feature differences between them and their neighboring nodes, and perform Byzantine consensus on clock deviation between nodes: If the difference between the relative clock deviation independently measured by more than a preset proportion of neighboring nodes and the absolute clock deviation conversion value reported by the node itself is greater than a preset tolerance, it is determined that there is timestamp tampering or clock synchronization attack, calculate the consistency score, and mark the node as a highly suspicious node when the consistency score is lower than a preset threshold. Level 3: For highly suspicious nodes, compare the changing trends of their thermal entropy growth rate with the total entropy growth rate of the cluster, and perform time-series causal relationship analysis on the thermal entropy growth rate and power consumption to determine whether there is sensor tampering or hidden computing tasks.

8. The security detection system for a server node as described in claim 7, characterized in that, The collaborative decision and response module also includes a random forest classifier. The inputs of the random forest classifier are the feature reconstruction error calculated at the first level, the consistency score calculated at the second level, and the thermodynamic deviation calculated at the third level. The outputs are threat type labels and confidence levels. The threat type labels include: normal, mining program, ransomware, DDoS botnet, firmware tampering, hardware trojan, side-channel eavesdropping, and clock synchronization attack.

9. The security detection system for a server node as described in claim 1, characterized in that, The response actions include: recording alarms, notifying the administrator, remotely restarting the node via BMC, physically powering off and isolating the node, and triggering memory dump for evidence collection; when the threat type is firmware tampering or hardware Trojan, the collaborative decision and response module automatically performs physical power-off and notifies the replacement of hardware; the collaborative decision and response module also retains the original sensor data and clock deviation matrix for a preset time before and after the alarm time as digital forensic evidence.

10. A security detection system for a server node as described in claim 1, characterized in that, In the physical field modeling and anomaly detection module, the time window length used to construct the node feature matrix is ​​a first preset duration, and the sliding step size is a second preset duration. The multimodal physical signal acquisition module, clock deviation measurement module, physical field modeling and anomaly detection module, and collaborative decision and response module exchange data through an out-of-band management network. All sensor data is timestamped by the node's BMC during acquisition. When the detection result is confirmed as a false alarm, the data within the false alarm time window is used as a normal sample, and the model parameters of the graph variational autoencoder are incrementally updated.

Images