A method and system for integrating a side-channel attack-resistant encryption operation in a DPU

Abstract

The application discloses a kind of integrated in DPU's side channel attack prevention encryption operation method and system, belong to computer chip security technical field, it includes: generating programmable perturbation instruction and injecting analog perturbation signal in response to encryption operation instruction, collection environment feedback signal and extract physical response characteristic quantity;Characteristic quantity is converted into dynamic defense parameter, control encryption engine loading parameter executes operation and synchronously collects side channel observation data;Based on observation data, evaluate defense effect and generate adaptive adjustment instruction, feedback adjustment perturbation signal injection characteristics and update historical evaluation data set.The application adopts dynamic perturbation based on physical environment feedback and AI closed-loop adaptive adjustment mechanism, can resist complex side channel analysis to DPU in real time, solve the defects that traditional mask scheme performance loss is big and strategy is fixed, improve the encryption operation security of data processing unit in high-speed network environment.

Description

Technical Field

[0001] This invention relates to the field of computer chip security technology, and in particular to a method and system for preventing side-channel attacks by integrating a DPU. Background Technology

[0002] Data processing units (DPUs), as a new generation of computing infrastructure, play a core role in high-performance computing and high-speed network data encryption and decryption scenarios. During encryption operations, the power consumption information, electromagnetic radiation, and other physical characteristics generated by the chip hardware circuits are often strongly correlated with the key data, making them vulnerable to side-channel attacks. Therefore, it is essential to construct effective security protection mechanisms.

[0003] In related technologies, Chinese invention patent CN112543094A discloses a method for implementing DES masking to resist side-channel attacks. In each round of DES operation, different random sources are used to generate different random numbers to participate in the operation to achieve the purpose of masking the intermediate value. Moreover, the intermediate value calculated in each round is different from the intermediate value of the original DES algorithm, thus achieving the purpose of masking the original intermediate value. In addition, the mask value of the previous round is eliminated in each round of operation, saving the complicated steps of demasking and saving resources.

[0004] Regarding the aforementioned technologies, the inventors believe that this solution relies on a logic-level masking algorithm, resulting in a relatively fixed defense strategy. Furthermore, the generation and injection of random numbers are disconnected from the chip's transient physical environment. Attackers can still build statistical models through long-term power consumption observations, leading to high predictability of the defense and difficulty in adapting to increasingly complex attack methods, resulting in a low security ceiling. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a side-channel attack prevention encryption computation method and system integrated into a DPU. It employs a closed-loop adaptive method that actively injects controllable disturbances into the chip's internal physical environment and utilizes their unique feedback to generate multi-dimensional defense parameters in real time. This method can construct an unpredictable and continuously self-optimizing endogenous security protection system.

[0006] To achieve the above objectives, this application adopts the following technical solution:

[0007] Firstly, a side-channel attack prevention encryption operation method integrated into a Digital Processing Unit (DPU) is provided, comprising: responding to an encryption operation start command, generating a programmable perturbation command, and sending the programmable perturbation command to a programmable perturbation generator with signal generation function; controlling the programmable perturbation generator to inject simulated perturbation signals into the internal physical environment of the DPU chip according to the programmable perturbation command, and simultaneously acquiring environmental feedback signals triggered by the simulated perturbation signals to obtain dynamic feedback data; performing hardware feature extraction on the dynamic feedback data to obtain several physical response feature quantities, and, according to mapping rules used to implement numerical transformation, converting the physical response feature quantities... The parameters are converted into dynamic defense parameters corresponding to this encryption operation; the encryption engine is controlled to load the dynamic defense parameters to execute the encryption operation, and side-channel observation data is collected simultaneously during the encryption operation; the defense effect evaluation result is derived based on the side-channel observation data, and an adaptive adjustment instruction is generated based on the defense effect evaluation result. The adaptive adjustment instruction is fed back to the generation process of the programmable perturbation instruction to adjust the injection characteristics of the simulated perturbation signal in subsequent encryption operation cycles. Simultaneously, the physical response characteristics, dynamic defense parameters, and side-channel observation data of this encryption operation are stored in a set of historical evaluation data to achieve continuous updating of the defense knowledge base.

[0008] Based on the above technical solution, in the anti-side-channel attack encryption calculation method integrated into the DPU provided in this application, a closed-loop adaptive method is adopted to actively inject controllable disturbances into the internal physical environment of the chip and use its unique feedback to generate multi-dimensional defense parameters in real time. This method can build an unpredictable and continuously self-optimizing endogenous security protection system.

[0009] In conjunction with the first aspect above, in one possible implementation, controlling the programmable disturbance generator to inject simulated disturbance signals into the internal physical environment of the DPU chip according to the programmable disturbance instructions, and simultaneously acquiring environmental feedback signals triggered by the simulated disturbance signals, includes: controlling the programmable disturbance generator to generate electromagnetic micro-disturbance signals or power ripple signals with specific frequencies and amplitudes as simulated disturbance signals according to the programmable disturbance instructions; coupling the simulated disturbance signals into the power supply network or physical links of the input / output interfaces of the DPU chip to interact with the chip's operating state; and simultaneously monitoring voltage, current, or temperature changes at the injection point and associated areas through several on-chip sensors deployed around the encryption computing unit to acquire environmental feedback signals.

[0010] In conjunction with the first aspect above, in one possible implementation, hardware feature extraction is performed on the dynamic feedback data to obtain several physical response feature quantities, including: identifying from the environmental feedback signal the delay time from the moment the simulated disturbance signal is injected until the voltage or current value of a monitoring point recovers to a preset stable threshold range; calculating the energy attenuation coefficient of the environmental feedback signal on a specific frequency component of the simulated disturbance signal, wherein the delay time and the energy attenuation coefficient constitute physical response feature quantities.

[0011] In conjunction with the first aspect above, in one possible implementation, converting the physical response feature quantity into dynamic defense parameters corresponding to this encryption operation includes: inputting the delay time into a hardware hash function used to implement nonlinear transformation to generate a first mask sequence; performing an association operation between the energy attenuation coefficient and an identifier pool containing several hardware unit identifiers to determine the target hardware computing unit number, wherein the dynamic defense parameters include the first mask sequence and the target hardware computing unit number.

[0012] In conjunction with the first aspect above, in one possible implementation, the control encryption engine loading the dynamic defense parameters to perform encryption operations includes: using the first mask sequence to perform masking and obfuscation operations on plaintext data or intermediate state data entering the encryption engine; scheduling the encryption operation task after masking and obfuscation operations to the corresponding physically isolated hardware computing unit inside the DPU for execution according to the target hardware computing unit number; and during the execution of operations by the hardware computing unit, acquiring its power consumption trajectory or electromagnetic radiation waveform through a high-precision analog-to-digital converter to obtain side-channel observation data.

[0013] In conjunction with the first aspect above, in one possible implementation, converting the physical response feature into the dynamic defense parameter corresponding to this encryption operation further includes: based on the physical response feature, addressing in a lookup table used to store phase adjustment values ​​to determine the clock phase fine-tuning amount; when the control encryption engine loads the dynamic defense parameter to perform the encryption operation, applying the clock phase fine-tuning amount during the critical clock cycle of the encryption engine's operation to generate controllable clock jitter.

[0014] In conjunction with the first aspect above, in one possible implementation, generating an adaptive adjustment command based on the defense effectiveness evaluation result includes: performing time-domain or frequency-domain feature analysis on the side-channel observation data to obtain the side-channel leakage index for the current encryption operation cycle; comparing the side-channel leakage index with a set of historical evaluation data, the set including a standard power consumption template of the encryption engine collected under a preset security environment, and task execution performance indicators of encryption operations under different perturbation modes; obtaining a defense effectiveness evaluation result by comparing the deviation between the current leakage index and the standard power consumption template, the defense effectiveness evaluation result indicating whether the current defense is effective or needs to be enhanced; when the defense effectiveness evaluation result indicates that the defense needs to be enhanced, generating an adaptive adjustment command to increase the amplitude of the simulated perturbation signal or change its frequency.

[0015] In conjunction with the first aspect above, in one possible implementation, the comparison of the side-channel leakage index with a set of historical evaluation data is performed by a lightweight reinforcement learning model deployed in the DPU security isolation zone. The lightweight reinforcement learning model receives several sets of historical physical response features, dynamic defense parameters, and corresponding side-channel leakage indices as training data, and its output is an optimization suggestion for the programmable perturbation command parameters. The optimization suggestion constitutes the core part of the adaptive adjustment command.

[0016] In conjunction with the first aspect above, in one possible implementation, the method further includes: receiving security level configuration information from an upper-layer application before generating the programmable perturbation instruction; selecting an initial perturbation mode template from a set storing several perturbation mode templates based on the security level configuration information; and determining the generated programmable perturbation instruction based on the initial perturbation mode template and the adaptive adjustment instruction.

[0017] Secondly, a side-channel attack prevention encryption computing system integrated into a DPU is provided, comprising: an instruction generation module, used to generate a programmable perturbation instruction in response to an encryption computing start instruction, and send the programmable perturbation instruction to a programmable perturbation generator with signal generation function; a physical perturbation and acquisition module, used to control the programmable perturbation generator to inject simulated perturbation signals into the internal physical environment of the DPU chip according to the programmable perturbation instruction, and simultaneously acquire environmental feedback signals triggered by the simulated perturbation signals to obtain dynamic feedback data; and a feature extraction and conversion module, used to perform hardware feature extraction on the dynamic feedback data to obtain several physical response feature quantities, and convert the physical response data into numerical transformation based on mapping rules used to achieve numerical transformation. The characteristic quantity should be converted into the dynamic defense parameters corresponding to this encryption operation; the encryption execution and observation module is used to control the encryption engine to load the dynamic defense parameters to execute the encryption operation, and simultaneously collect side channel observation data during the encryption operation process; the defense evaluation and feedback storage module is used to derive the defense effect evaluation result based on the side channel observation data, and generate an adaptive adjustment instruction based on the defense effect evaluation result, feed the adaptive adjustment instruction back to the generation process of the programmable perturbation instruction, adjust the injection characteristics of the simulated perturbation signal in the subsequent encryption operation cycle, and simultaneously store the physical response characteristic quantity, dynamic defense parameters and side channel observation data of this encryption operation into a set of historical evaluation data to realize the continuous updating of the defense knowledge base.

[0018] Compared with the prior art, the present invention has the following advantages:

[0019] This invention eliminates the predictability of defensive actions by constructing a dynamic closed loop of active detection and environmental feedback. Since the random parameters used in each defense are endogenously derived from the unique interaction between the system's own active behavior and the unpredictable instantaneous physical environment, attackers cannot build effective attack models through long-term observation, thus constructing an inherent immunity to advanced side-channel attack methods based on statistical analysis and machine learning.

[0020] This invention achieves deep synergy among the data layer, logic layer, and temporal layer defense mechanisms. Feature quantities extracted from feedback from a single physical environment are simultaneously used to drive the generation of data masks, the physical relocation of computational units, and the phase jitter of critical clocks. This multi-dimensional, homogeneous defense strategy ensures that the various defense measures are interconnected and change synchronously, forming a comprehensive protection effect that surpasses the simple superposition of independent applications.

[0021] This invention establishes an adaptive learning and control mechanism based on actual defense effectiveness. It can quantify the effectiveness of its own defensive behavior in real time and use the evaluation result as a negative feedback signal to continuously optimize subsequent defense strategies. This gives the defense system the ability to evolve and self-reinforce over the long term, automatically adapting to chip aging, environmental changes, and constantly evolving attack methods, ensuring the long-term robustness of security protection.

[0022] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This application provides a structural architecture diagram of a side-channel attack prevention encryption computing system integrated into a DPU, as shown in the embodiments of this application.

[0025] Figure 2 A flowchart illustrating a side-channel attack prevention encryption computation method integrated into a DPU, provided for an embodiment of this application;

[0026] Figure 3 This is a transient response curve of the environmental feedback signal provided in the embodiments of this application.

[0027] Figure 4 This is a graph showing the change in the side channel leakage index during the dynamic adjustment process provided in this application embodiment. Detailed Implementation

[0028] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0029] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0030] The side-channel attack prevention encryption method integrated into a DPU provided in this application embodiment can be applied to, for example... Figure 1 The illustrated example is a side-channel attack prevention encryption computing system 100 integrated into a DPU, such as... Figure 1 As shown, the system includes:

[0031] The instruction generation module is used to generate a programmable perturbation instruction in response to the encryption operation start instruction, and send the programmable perturbation instruction to a programmable perturbation generator with signal generation function;

[0032] The physical perturbation and acquisition module is used to control the programmable perturbation generator to inject simulated perturbation signals into the internal physical environment of the DPU chip according to the programmable perturbation instructions, and simultaneously acquire the environmental feedback signals triggered by the simulated perturbation signals to obtain dynamic feedback data.

[0033] The feature extraction and conversion module is used to perform hardware feature extraction on the dynamic feedback data to obtain several physical response feature quantities, and convert the physical response feature quantities into dynamic defense parameters corresponding to this encryption operation according to the mapping rules used to realize numerical transformation.

[0034] The encryption execution and observation module is used to control the encryption engine to load the dynamic defense parameters to perform encryption operations, and to synchronously collect side-channel observation data during the encryption operation process.

[0035] The defense assessment and feedback storage module is used to derive the defense effectiveness assessment result based on the side channel observation data, generate an adaptive adjustment instruction based on the defense effectiveness assessment result, feed the adaptive adjustment instruction back to the generation process of the programmable perturbation instruction, adjust the injection characteristics of the simulated perturbation signal in the subsequent encryption operation cycle, and simultaneously store the physical response characteristics, dynamic defense parameters and side channel observation data of this encryption operation into a set of historical assessment data to realize the continuous updating of the defense knowledge base.

[0036] like Figure 2 As shown in the embodiment of this application, a side-channel attack prevention encryption computation method integrated into a DPU is provided, including:

[0037] In response to the encryption operation start command, a programmable perturbation command is generated and sent to a programmable perturbation generator with signal generation function;

[0038] The programmable disturbance generator is controlled to inject simulated disturbance signals into the internal physical environment of the DPU chip according to the programmable disturbance instructions, and simultaneously collects environmental feedback signals triggered by the simulated disturbance signals to obtain dynamic feedback data;

[0039] Hardware feature extraction is performed on the dynamic feedback data to obtain several physical response feature quantities, and the physical response feature quantities are converted into dynamic defense parameters corresponding to this encryption operation according to the mapping rules used to realize numerical transformation.

[0040] The encryption engine is controlled to load the dynamic defense parameters to perform encryption operations, and side-channel observation data is collected synchronously during the encryption operation process;

[0041] Based on the side-channel observation data, the defense effectiveness evaluation result is derived, and an adaptive adjustment command is generated according to the defense effectiveness evaluation result. The adaptive adjustment command is fed back to the generation process of the programmable perturbation command to adjust the injection characteristics of the simulated perturbation signal in the subsequent encryption operation cycle. At the same time, the physical response characteristics, dynamic defense parameters and side-channel observation data of this encryption operation are stored in the set of historical evaluation data to realize the continuous updating of the defense knowledge base.

[0042] It is important to note that a physical feedback-based adaptive closed-loop defense system is constructed. Its core involves actively injecting controllable perturbation signals into the internal physical environment of the DPU chip as excitation, and capturing in real time the transient response triggered by this excitation, which reflects the chip's current unique physical state. This method transforms this simulated, unpredictable physical response into a set of dynamic defense parameters for the current encryption operation through hardware feature extraction and deterministic numerical transformation rules. These parameters are applied to the encryption engine in real time to obfuscate its operation. Simultaneously, real-time side-channel leakage observation and evaluation are performed on the encryption process after the application of defense measures, and the evaluation results are used as feedback signals to dynamically adjust the characteristics of the perturbation signals injected in the next encryption cycle. Through continuous perturbation, acquisition, transformation, application, evaluation, and feedback, and by continuously accumulating defense process data to form a knowledge base, this method establishes an intelligent defense mechanism that can bind the chip's physical uniqueness to the encryption defense strength in real time and can autonomously optimize based on the actual defense effect.

[0043] In one possible implementation of the embodiments of this application, combined with Figure 2 The programmable disturbance generator is controlled to inject simulated disturbance signals into the internal physical environment of the DPU chip according to the programmable disturbance instructions, and simultaneously collects environmental feedback signals triggered by the simulated disturbance signals, including:

[0044] The programmable disturbance generator is controlled to generate an electromagnetic disturbance signal or a power ripple signal with a specific frequency and amplitude according to the programmable disturbance instruction, as a simulated disturbance signal;

[0045] The simulated disturbance signal is coupled and injected into the power supply network or physical link of the input / output interface of the DPU chip to interact with the chip's operating state.

[0046] By deploying several on-chip sensors around the encryption computing unit, the voltage, current, or temperature changes of the injection point and associated area are monitored synchronously to obtain environmental feedback signals.

[0047] In some implementations, upon receiving a programmable perturbation command carrying specific parameters, the programmable perturbation generator is first controlled to generate a simulated perturbation signal based on the signal type, frequency, and amplitude parameters parsed from the command. This signal can be a power supply ripple signal with a frequency range between tens of kilohertz and hundreds of megahertz, and its amplitude is typically set to 0.5% to 5% of the DPU core power supply voltage to ensure sufficient measurable perturbation without affecting the chip's normal logic function. Alternatively, the signal can be an electromagnetic perturbation signal in a specific frequency band, directionally coupled to sensitive areas via a miniature antenna. Subsequently, this simulated perturbation signal is precisely injected into the physical environment of the DPU chip. The injection path is selected according to the command and can be the chip's power supply network, for example, directly coupled to one or more nodes of the on-chip power distribution network; or it can be the "physical link of the input / output interface," such as introducing common-mode perturbation on the differential traces of a high-speed serial interface. The purpose of the injection is to allow this perturbation to interact physically with the power dynamics generated by the encryption computing unit during task execution. Simultaneously with the injection of a simulated disturbance signal, multiple on-chip sensors deployed around the encryption processing unit are activated. These sensor arrays consist of high-bandwidth voltage probes, current sensors, or thermistors, which synchronously start monitoring and collect instantaneous changes in voltage, current, or temperature at the injection point and surrounding key areas in real time. This series of continuous sampled values ​​constitutes a multi-channel environmental feedback signal, with a time-series data format that accurately records the entire dynamic process of the chip's physical environment from receiving the disturbance to returning to a steady state. As shown in Figure 3, the transient response curve of the environmental feedback signal of the DPU chip after the injection of the simulated disturbance signal in this embodiment of the invention is shown. This curve reflects the physical dynamic process of the voltage gradually converging to a preset stable threshold range after the disturbance fluctuation, and is the original data basis for extracting features such as delay time.

[0048] For example, with encryption operation as the trigger condition, the programmable perturbation generator, based on the issued instruction, analyzes the signal type as power ripple, frequency as 100MHz, and amplitude as 2%. If the current DPU core power supply voltage is 1.0V, the generator produces an amplitude of 0.02V, calculated as follows: The simulated disturbance signal is injected directly into the critical nodes of the on-chip power distribution network via a coupling circuit. This disturbance signal physically interferes with the power consumption waveform of the encryption operation unit in high-frequency switching mode. At this time, the on-chip voltage probe array distributed around the operation unit is synchronously activated at a sampling rate of 10 GHz, capturing the tiny instantaneous fluctuations of voltage between 0.98V and 1.02V in real time. Assuming that the sensor continuously collects 1000 sample points within a 100ns window of the disturbance occurrence, these time-series data recording the entire process of voltage from the disturbed fluctuation to convergence to the steady-state threshold constitute the multi-channel environmental feedback signal, providing raw dynamic feedback data for subsequent extraction of the chip's unique physical characteristics.

[0049] In one possible implementation, combining Figure 2 Hardware feature extraction is performed on the dynamic feedback data to obtain several physical response feature quantities, including:

[0050] From the environmental feedback signal, identify the delay time from the moment the simulated disturbance signal is injected until the voltage or current value at a monitoring point recovers to a preset stable threshold range;

[0051] Calculate the energy attenuation coefficient of the environmental feedback signal at a specific frequency component of the simulated disturbance signal, and the delay time and the energy attenuation coefficient constitute a physical response characteristic quantity.

[0052] In some implementations, this process is executed directly at the hardware level, with the input being time-series data of the environmental feedback signal. Two core metrics are calculated in parallel. The first metric is latency. The hardware logic unit starts a picosecond-precision time counter from the precise time point of the simulated disturbance signal injection. Simultaneously, it continuously monitors the output signal of specific on-chip sensors, such as the voltage value on the power rail of the encryption engine core. This voltage value is compared in real time with a preset stability threshold range, typically defined as the interval between ±0.1% and ±1% of the nominal operating voltage. The time counter stops when the monitored voltage value returns to and remains stable within this threshold range for more than a preset debouncing time after the disturbance. The final reading of the counter is the latency, which physically reflects the transient response characteristics of the chip's power distribution network. The second metric is the energy attenuation coefficient. The hardware performs a Fast Fourier Transform on a specific time window of data from the captured environmental feedback signal to analyze its spectrum. This transform is also performed by a dedicated hardware unit to meet real-time requirements. The energy attenuation coefficient is used to quantify the ability of the DPU chip's physical structure to absorb or dissipate energy from disturbances at a specific frequency. Its calculation follows the formula:

[0053] ;

[0054] in, It represents the energy decay coefficient and is a dimensionless scalar. This represents the power spectral density value of the environmental feedback signal collected at the characteristic frequency point of the injected simulated disturbance signal. This value is obtained by extracting the energy value of the corresponding frequency bin after hardware FFT calculation. This represents the reference power spectral density value of the injected simulated perturbation signal at this characteristic frequency point. This value is obtained as a known parameter from the programmable perturbation generator. The calculated delay time and energy attenuation coefficient together constitute the physical response characteristic quantity of this encryption operation cycle and are immediately sent to the subsequent processing unit.

[0055] For example, the hardware logic performs parallel processing on the acquired time-series environmental feedback signals to extract feature quantities. For the calculation of the delay time, a picosecond-level counter is synchronously started at the beginning of the simulated disturbance signal injection, and the nominal voltage of the encryption engine power rail (1.0V) is monitored in real time. If the preset stability threshold range is ±0.5%, i.e., 0.995V to 1.005V, and the monitored voltage returns to this threshold range for the first time after 135 picoseconds following the disturbance and meets the preset de-jitter time requirement, then the determined delay time feature value is 135 picoseconds. This value physically reflects the transient response characteristics of the chip's power distribution network. Simultaneously, the hardware computing unit performs a Fast Fourier Transform on a specific time window of data from the environmental feedback signal, analyzes its spectrum, and extracts the energy distribution at the disturbance feature frequency point. If the reference power spectral density of the injected signal at this feature frequency point is known... The value is 100 units, while the measured power spectral density of the feedback signal at that point is... If the value is 72 units, then according to the formula... The energy attenuation coefficient of the physical structure of the quantized chip was calculated. The value is 0.72. The calculation process is as follows: Ultimately, the physical response characteristics of this encryption cycle, consisting of a delay time of 135 picoseconds and an energy decay coefficient of 0.72, serve as a quantified and physically unique input for the subsequent defense parameter generation stage.

[0056] In one possible implementation, combining Figure 2 Converting the physical response features into dynamic defense parameters corresponding to this encryption operation includes:

[0057] The delay time is input into a hardware hash function to implement a nonlinear transformation, generating a first mask sequence;

[0058] The energy attenuation coefficient is associated with an identifier pool containing several hardware unit identifiers to determine the target hardware computing unit number. The dynamic defense parameters include the first mask sequence and the target hardware computing unit number.

[0059] In some implementations, this conversion process is implemented using dedicated hardware logic that takes the delay time and energy attenuation coefficient as inputs and generates two independent defense parameters in parallel. First, to generate the first mask sequence, the measured delay time is used as an entropy source and directly input into a built-in hardware hash function. This hash function is a pure combinational logic circuit capable of performing a high avalanche effect nonlinear transformation within a single clock cycle. Its output bit width matches the data block width of the encryption algorithm; for example, for the AES algorithm, the output is a 128-bit sequence. This output sequence is the first mask sequence. Due to the one-way nature of the hash function, even if an attacker knows the mask sequence, they cannot deduce the physical characteristic of the delay time. Second, to determine the target hardware computing unit number, the dimensionless energy attenuation coefficient is used for an association operation. This operation is essentially an index mapping process, selecting the target unit from a pool of identifiers that pre-stores multiple hardware unit identifiers. This selection process follows the following operations:

[0060] ;

[0061] in, It is the calculated target index value. It is the input energy attenuation coefficient. It is the total number of available hardware computing units recorded in the identifier pool, which is a fixed system configuration parameter. It is a non-linear scaling function implemented in hardware, designed to... The floating-point values ​​are mapped to a sufficiently large integer space to ensure uniformity of selection; for example, it can be implemented as follows: ,in It is a fixed, large integer scaling factor. The hardware logic scales based on the calculated index. The corresponding hardware unit identifier is retrieved from the identifier pool; this identifier is the target hardware computing unit number. The generated first mask sequence and the target hardware computing unit number are encapsulated into dynamic defense parameters.

[0062] For example, the hardware logic receives the extracted physical response features, including a delay time of 135 picoseconds and an energy attenuation coefficient of 0.72, and converts them into dynamic defense parameters to guide the encryption process. First, the delay time value of 135 is input into a built-in hardware hash function. Through the nonlinear transformation of this combinational logic circuit, a 128-bit first mask sequence matching the encryption algorithm is generated, ensuring an irreversible mapping from physical features to the mask space. Simultaneously, the energy attenuation coefficient of 0.72 is applied to the index mapping operation to determine execution resources, such as a fixed scaling factor. The total number of available hardware computing units is 1000. If it is 8, then according to the formula The calculation is performed by first obtaining the integer value 720 through a non-linear scaling function. The calculation process is as follows: Then, a modulo operation is performed to obtain the target index value 0. The calculation process is as follows: The hardware logic then retrieves the hardware unit identifier corresponding to index 0 from the identifier pool, thus determining the target hardware computing unit number as "Unit_0". Finally, the generated 128-bit first mask sequence and the target hardware computing unit number "Unit_0" are encapsulated as dynamic defense parameters for this encryption cycle, thereby achieving randomized hardening of the encryption operation in two dimensions: data obfuscation and hardware resource scheduling.

[0063] In one possible implementation, combining Figure 2 The control encryption engine loading the dynamic defense parameters to perform encryption operations includes:

[0064] The first mask sequence is used to perform masking and obfuscation operations on the plaintext data or intermediate state data entering the encryption engine;

[0065] Based on the target hardware computing unit number, the encryption operation task after masking and obfuscation is scheduled to be executed on the corresponding physically isolated hardware computing unit inside the DPU.

[0066] During the execution of operations by the hardware computing unit, its power consumption trajectory or electromagnetic radiation waveform is collected by a high-precision analog-to-digital converter to obtain side-channel observation data.

[0067] In some implementations, the process is triggered by received dynamic defense parameters. The masking logic unit performs a bit-by-bit XOR operation between the first mask sequence in the parameters and the plaintext data about to enter the encryption engine's processing pipeline, or the intermediate state data flowing between encryption round functions, using a hardware XOR gate. This obfuscation operation aims to sever the direct correlation between encrypted data and power leakage. Simultaneously, the target hardware computing unit number in the parameters is parsed, and an on-chip scheduling network, such as a hardware multiplexer or routing switch unit, is configured accordingly. This scheduling network precisely assigns the obfuscated encryption task, including the masked data and related control signals, to a specific, physically isolated hardware computing unit within the DPU chip, with its own independent power supply and clock domain. These units possess different side-channel characteristics due to their different positions on the chip layout, trace lengths, and process variations. At the instant the designated hardware computing unit begins executing the encryption operation, the output signal of the sensor tightly coupled to that unit is sampled at high speed using a high-precision analog-to-digital converter. The sampling target can be the voltage drop across a tiny resistor connected in series in the unit's power supply network to obtain the power consumption trajectory; or it can be the electromagnetic field strength sensed by a miniature antenna placed above the unit to obtain the electromagnetic radiation waveform. The analog-to-digital converter typically uses a sampling rate in the gigahertz range and a quantization precision of 8 to 12 bits to ensure the capture of weak signals associated with a single gate flip. The resulting sequence of digitized sample points constitutes the side-channel observation data for this encryption operation and is temporarily stored for subsequent analysis.

[0068] For example, upon receiving dynamic defense parameters containing a 128-bit first mask sequence and the target hardware computing unit number "Unit_0", the hardware logic immediately triggers a dual randomization defense operation. First, the mask logic unit performs a bit-by-bit XOR operation on the 128-bit plaintext data and the first mask sequence using a hardware XOR gate, severing the direct link between plaintext and power leakage through data-level obfuscation. Subsequently, the scheduling control unit parses the number "Unit_0" and configures the on-chip routing and switching unit to schedule the masked task to the "Unit_0" unit, which has independent power supply and clock domains, for execution, utilizing its unique physical routing and layout to achieve spatial-level position obfuscation. At the instant "Unit_0" executes its operation, a high-precision analog-to-digital converter with a sampling rate of 10GHz and a quantization precision of 12 bits is activated. The voltage drop across a 10 milliohm precision resistor connected in series on the "Unit_0" power supply rail is sampled. If the measured instantaneous voltage drop is 5 millivolts, the instantaneous power consumption current is calculated to be 0.5 amps according to Ohm's law. The calculation process is as follows: The digitized sequence of thousands of such current sample points generated by this operation constitutes the side-channel observation data, which is temporarily stored in a secure storage area for subsequent defense effectiveness evaluation.

[0069] In one possible implementation, combining Figure 2 The conversion of the physical response features into dynamic defense parameters corresponding to this encryption operation also includes:

[0070] Based on the physical response characteristics, the clock phase fine-tuning amount is determined by addressing in a lookup table used to store phase adjustment values.

[0071] When the control encryption engine loads the dynamic defense parameters to perform encryption operations, the clock phase fine-tuning amount is applied during the critical clock cycle of the encryption engine's operation to generate controllable clock jitter.

[0072] In some implementations, this process is performed in parallel with masking and cell selection, supplementing the generation of dynamic defense parameters. After obtaining the physical response characteristics, an additional mapping operation is performed. Using these characteristics, such as delay time and energy decay coefficient, an address is generated for addressing in a pre-defined lookup table storing phase adjustment values. This lookup table is essentially a small read-only memory whose contents are fixed after chip manufacturing, where each address corresponds to a specific digital code used to control the clock phase. Address generation follows the following operations:

[0073] ;

[0074] in, It is the address of the generated lookup table. This represents a dedicated, low-latency hardware hash function used to uniformly map input data across the entire address space. The binary representation of the delay time. This is the binary representation of the energy decay coefficient after fixed-point normalization. Symbol This indicates that two binary sequences are concatenated to form a longer input bit string, thereby increasing the input entropy of the hash function. Through this address... The value read from the lookup table is the clock phase adjustment for this encryption operation, which physically corresponds to a picosecond-level delay or advance. Subsequently, during the execution of the encryption engine's operation, its internal timing control unit loads this clock phase adjustment into the numerically controlled delay line of the clock management unit during specific critical clock cycles, such as the cycles for performing nonlinear substitution boxes or modular multiplication operations. This causes a tiny, physically uniquely determined offset of the rising edge of the clock signal in that specific clock cycle relative to its nominal position, thereby introducing controllable clock jitter on the timeline of the encryption operation, effectively disrupting the attacker's foundation for power trace alignment and averaging.

[0075] For example, after acquiring physical response characteristics, the hardware logic further strengthens the defense by introducing randomization in the time dimension. This is based on the measured latency. 135 picoseconds, represented in binary as 10000111 and energy decay coefficient. Taking a value of 0.72 and a fixed-point binary representation of 101110 as an example, first, the two binary sequences are concatenated to form the input bit string 10000111101110, and then input into a low-latency hardware hash function. According to the formula Assuming the hash function maps the input to lookup table address 0x3F, and then reads the phase code value corresponding to that address from a preset read-only memory, if the physical offset corresponding to this code value is 5 picoseconds, then the clock phase fine-tuning amount for this operation is determined to be 5 picoseconds. During the critical clock cycle of the encryption engine's nonlinear substitution box operation, the timing control unit loads this 5 picosecond offset into the numerically controlled delay line, causing the rising edge of the clock signal in that cycle to have a 5 picosecond physical offset relative to its nominal position. This clock jitter, uniquely determined by the physical response, causes a nonlinear displacement of the encryption operation on the time axis, effectively disrupting the possibility of a side-channel attacker performing power trace time alignment.

[0076] In one possible implementation, combining Figure 2 The adaptive adjustment instructions generated based on the defense effectiveness evaluation results include:

[0077] Perform time-domain or frequency-domain feature analysis on the side-channel observation data to obtain the side-channel leakage index for the current encryption operation cycle;

[0078] The side-channel leakage index is compared with a set of historical evaluation data, which includes a standard power consumption template of the encryption engine collected under a preset security environment, and task execution performance indicators of encryption operations under different perturbation modes. By comparing the deviation between the current leakage index and the standard power consumption template, the defense effectiveness evaluation result is obtained. The defense effectiveness evaluation result indicates whether the current defense is effective or the defense needs to be strengthened.

[0079] When the defense effectiveness evaluation results indicate that the defense needs to be enhanced, an adaptive adjustment command is generated to increase the amplitude of the simulated disturbance signal or change its frequency.

[0080] In some implementations, the side-channel observation data collected during the encryption operation is input. First, a dedicated analysis engine performs time-domain or frequency-domain feature analysis on the side-channel observation data to calculate a quantified side-channel leakage index. The highest correlation peak in the correlation power consumption analysis directly reflects the correlation strength between power consumption and the intermediate value of key correlation. This index can be expressed as:

[0081] ;

[0082] in, This is the side-channel leakage index, a dimensionless value. The higher the value, the more severe the leakage. This represents the calculation operation for the Pearson correlation coefficient. It is the time series of the input side-channel observation data. It is a set of hypothetical models, generated by the hardware based on the structure of the encryption algorithm and all possible values ​​of intermediate values ​​related to a specific key, resulting in a set of hypothetical power consumption trajectories. The hardware will quickly calculate... and The correlation of each trajectory is analyzed, and the maximum value is taken as the correlation. Subsequently, the calculated side-channel leakage index was... It is compared to a preset security threshold stored in a set of historical evaluation data. This set, the defense knowledge base, also includes pre-collected standard power consumption templates for encryption engines under undisturbed, absolutely secure environments, as well as performance metrics for encryption tasks, such as computational latency, measured under different perturbation modes. Current Leakage Index The deviation from the standard power consumption template is essentially The value itself, compared with the security threshold, constitutes the defense effectiveness assessment result. If If the threshold is exceeded, the assessment result indicates that defense needs to be enhanced. When the assessment result indicates that defense needs to be enhanced, an adaptive adjustment command is immediately generated. This command is a set of digital codes that clearly indicates what adjustments need to be made to the parameters of the simulated disturbance signal in the next cycle, such as increasing the signal amplitude by a preset step or switching its frequency to a backup frequency, in order to reduce side-channel leakage through stronger physical environment obfuscation. Figure 4 As shown, the side-channel leakage index is displayed during multiple consecutive adaptive adjustment cycles of encryption operations. The evolution trend of gradually decreasing and converging below the preset safety threshold as the closed-loop feedback adjustment mechanism is followed reflects the self-optimization and reinforcement capabilities of the defense system of this invention.

[0083] For example, through power consumption trajectory data generated from a single encryption task Feature analysis is performed to quantify the effectiveness of online defense. First, the hardware logic uses the measured power consumption sequence... With the set of hypothetical power consumption models generated for intermediate values ​​of encryption algorithms Perform the Pearson correlation coefficient calculation; if the calculated maximum correlation coefficient is 0.25, then follow the formula... Determine the side-channel leakage index for this encryption cycle. The value is 0.25. This index is then compared to a preset security threshold, such as 0.15, stored in the historical evaluation dataset. Since the calculated value of 0.25 is greater than the threshold of 0.15, the current defense is deemed ineffective and needs strengthening. Based on this determination, an adaptive adjustment instruction composed of digital codes is immediately generated, explicitly instructing the adjustment of the injection characteristics of the simulated perturbation signal in the next encryption cycle. For example, the amplitude ratio of the power supply ripple signal is increased from 2% of the supply voltage by a preset step, such as from 0.5% to 2.5%, to reduce the side-channel information leakage intensity in subsequent calculations through stronger physical environment obfuscation.

[0084] In one possible implementation, combining Figure 2 The comparison of the side-channel leakage index with a set of historical evaluation data is performed by a lightweight reinforcement learning model deployed in the DPU security isolation zone. The lightweight reinforcement learning model receives several sets of historical physical response features, dynamic defense parameters and corresponding side-channel leakage indices as training data, and its output is an optimization suggestion for the parameters of the programmable perturbation command. The optimization suggestion constitutes the core part of the adaptive adjustment command.

[0085] In some implementations, this intelligent decision-making process is executed by a lightweight reinforcement learning model deployed within a secure isolation zone on the DPU chip. This secure isolation zone is a dedicated processing environment with hardware access control and encrypted memory protection, ensuring the integrity of the model and the confidentiality of the decision-making process. The model is trained to understand the causal chain between perturbation, response, and leakage. At the end of each encryption cycle, the model receives the latest set of data as input for its continued learning. This data includes physical response features from the previous cycle, dynamic defense parameters generated based on these features, and the final evaluated side-channel leakage index. The model treats these three as feedback signals for state, action, and reward in the reinforcement learning paradigm, respectively. The model's internal policy function aims to learn an optimal mapping, i.e., what action should be taken to maximize future cumulative rewards, i.e., minimize long-term side-channel leakage, given a system state. When adaptive adjustment instructions need to be generated for the next encryption cycle, the model performs an inference operation, the output of which is an optimized suggestion for the parameters in the programmable perturbation instructions. This process is represented as:

[0086] ;

[0087] in, The next set of optimal perturbation parameters, representing the model output, is the optimization suggestion, which includes specific numerical adjustment suggestions for the amplitude and frequency of the simulated perturbation signal. This represents the policy function learned by the reinforcement learning model after training. This function may exist in the hardware in the form of neural network weights or a Q-table. This represents the current state, specifically the set of physical response features most recently acquired during this encryption cycle. This optimization suggestion is then encapsulated, forming the core data part of the adaptive adjustment instruction, and fed back to guide the generation of the next round of perturbation signals.

[0088] For example, a lightweight reinforcement learning model deployed within the secure isolation zone of the DPU chip achieves intelligent control through continuous learning of the defense process. At the end of the current encryption cycle, the model receives a current state set consisting of a 135 picosecond delay and an energy decay coefficient of 0.72. This is combined with the actions of the previous cycle, such as the injection magnitude of 2% and the reward feedback obtained, such as the side-channel leakage index of 0.25, to formulate an internal strategy function. The inference calculation. If the trained policy function determines that, under the current physical response state, increasing the perturbation frequency is more effective in suppressing leakage than simply increasing the amplitude, then according to the formula... Calculation yields optimization suggestions for the next cycle. Specifically, it is recommended to change the frequency of the simulated disturbance signal from... Adjust to The amplitude ratio was then fine-tuned to 2.25%. This optimization suggestion, which includes numerical adjustment recommendations, was then encapsulated as an adaptive adjustment instruction and fed back to the top layer. This leverages the decision-making capabilities of reinforcement learning to precisely intervene in physical perturbation characteristics in subsequent cycles, achieving self-evolving defense against complex nonlinear leakage scenarios.

[0089] In one possible implementation, combining Figure 2 The method further includes:

[0090] Before generating the programmable perturbation instruction, security level configuration information is received from the upper-layer application;

[0091] Based on the security level configuration information, an initial perturbation mode template is selected from a set of stored perturbation mode templates;

[0092] The programmable disturbance instruction is generated based on the initial disturbance mode template and the adaptive adjustment instruction.

[0093] In some implementations, this method introduces a policy setting step at the very beginning of the entire encryption defense cycle. Upon responding to the "encryption operation start command," it first receives security level configuration information from the upper-layer application or system management interface. This information is a discrete identifier, for example, indicating low, medium, and high security levels. Based on the received level identifier, a fixed storage area is accessed, containing a predefined set of perturbation pattern templates. Each template corresponds to a security level and specifies in detail the baseline parameters of the simulated perturbation signal. For example, a high-level template might specify high-amplitude broadband noise or multi-frequency combination perturbations, while a low-level template corresponds to low-amplitude single-frequency sinusoidal perturbations. An initial perturbation pattern template matching the current security level is selected through a lookup table operation. Finally, when generating the final programmable perturbation command, a fusion decision process is executed, integrating the baseline parameters provided by this initial perturbation pattern template with optimization suggestions contained in the adaptive adjustment command from the previous cycle, fed back from the defense knowledge base. For example, if the amplitude defined by the initial template is a baseline value... The adaptive instruction suggestion is to increase by 10%. The magnitude in the final instruction will then be set to the sum of the two values ​​or a new value calculated according to a specific fusion algorithm. In this way, both the macro-level security strategy at the upper level and the micro-level adjustments at the lower level based on actual attack and defense effects are adopted, achieving precise defense that combines strategy-driven and data-driven approaches.

[0094] For example, a layered control framework is used to combine pre-configuration and dynamic optimization of defense strength. When an encryption operation start command is received, security level configuration information is first obtained from the system management interface. If the information indicates a high security level, the corresponding baseline parameters are matched from a pre-set set of disturbance mode templates through a lookup table operation, such as setting the reference amplitude ratio of the simulated disturbance signal. 3% of the supply voltage, with a reference frequency of Subsequently, during the fusion decision process for generating the final programmable perturbation command, the reference amplitude is... Integrate with optimization suggestions from the previous cycle's adaptive adjustment instructions fed back from the defense knowledge base; if the optimization suggestions indicate a high risk of leakage in the current environment, it is recommended to increase the increment based on the baseline. If the value is 0.5%, then according to the fusion operation rule, the final injected signal amplitude ratio is calculated to be 3.5%. The calculation process is as follows: By setting this strategy, while ensuring adherence to the high-strength baseline of 3% set by the upper-level macro-security strategy, a micro-adjustment of +0.5% based on actual leakage assessment at the lower level is adopted, achieving precise security protection through a deep integration of strategy-driven and data-driven approaches.

[0095] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0096] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A side-channel attack prevention encryption computation method integrated into a DPU, characterized in that, The method includes: In response to the encryption operation start command, a programmable perturbation command is generated and sent to a programmable perturbation generator with signal generation function; The programmable disturbance generator is controlled to inject simulated disturbance signals into the internal physical environment of the DPU chip according to the programmable disturbance instructions, and simultaneously collects environmental feedback signals triggered by the simulated disturbance signals to obtain dynamic feedback data; Hardware feature extraction is performed on the dynamic feedback data to obtain several physical response feature quantities, and the physical response feature quantities are converted into dynamic defense parameters corresponding to this encryption operation according to the mapping rules used to realize numerical transformation. The encryption engine is controlled to load the dynamic defense parameters to perform encryption operations, and side-channel observation data is collected synchronously during the encryption operation process; Based on the side-channel observation data, the defense effectiveness evaluation result is derived, and an adaptive adjustment command is generated according to the defense effectiveness evaluation result. The adaptive adjustment command is fed back to the generation process of the programmable perturbation command to adjust the injection characteristics of the simulated perturbation signal in the subsequent encryption operation cycle. At the same time, the physical response characteristics, dynamic defense parameters and side-channel observation data of this encryption operation are stored in the set of historical evaluation data to realize the continuous updating of the defense knowledge base.

2. The anti-side-channel attack encryption computation method integrated into a DPU according to claim 1, characterized in that, The programmable disturbance generator is controlled to inject simulated disturbance signals into the internal physical environment of the DPU chip according to the programmable disturbance instructions, and the environmental feedback signals triggered by the simulated disturbance signals are collected simultaneously, including: The programmable disturbance generator is controlled to generate an electromagnetic disturbance signal or a power ripple signal with a specific frequency and amplitude according to the programmable disturbance instruction, as a simulated disturbance signal; The simulated disturbance signal is coupled and injected into the power supply network or physical link of the input / output interface of the DPU chip to interact with the chip's operating state. By deploying several on-chip sensors around the encryption computing unit, the voltage, current, or temperature changes of the injection point and associated area are monitored synchronously to obtain environmental feedback signals.

3. The anti-side-channel attack encryption computation method integrated into a DPU according to claim 1, characterized in that, Hardware feature extraction is performed on the dynamic feedback data to obtain several physical response feature quantities, including: From the environmental feedback signal, identify the delay time from the moment the simulated disturbance signal is injected until the voltage or current value at a monitoring point recovers to a preset stable threshold range; Calculate the energy attenuation coefficient of the environmental feedback signal at a specific frequency component of the simulated disturbance signal, and the delay time and the energy attenuation coefficient constitute a physical response characteristic quantity.

4. The anti-side-channel attack encryption computation method integrated into a DPU according to claim 3, characterized in that, Converting the physical response features into dynamic defense parameters corresponding to this encryption operation includes: The delay time is input into a hardware hash function to implement a nonlinear transformation, generating a first mask sequence; The energy attenuation coefficient is associated with an identifier pool containing several hardware unit identifiers to determine the target hardware computing unit number. The dynamic defense parameters include the first mask sequence and the target hardware computing unit number.

5. The anti-side-channel attack encryption computation method integrated into a DPU according to claim 4, characterized in that, The control encryption engine loading the dynamic defense parameters to perform encryption operations includes: The first mask sequence is used to perform masking and obfuscation operations on the plaintext data or intermediate state data entering the encryption engine; Based on the target hardware computing unit number, the encryption operation task after masking and obfuscation is scheduled to be executed on the corresponding physically isolated hardware computing unit inside the DPU. During the execution of operations by the hardware computing unit, its power consumption trajectory or electromagnetic radiation waveform is collected by a high-precision analog-to-digital converter to obtain side-channel observation data.

6. The anti-side-channel attack encryption computation method integrated into a DPU according to claim 4, characterized in that, Converting the physical response feature into dynamic defense parameters corresponding to this encryption operation also includes: Based on the physical response characteristics, the clock phase fine-tuning amount is determined by addressing in a lookup table used to store phase adjustment values. When the control encryption engine loads the dynamic defense parameters to perform encryption operations, the clock phase fine-tuning amount is applied during the critical clock cycle of the encryption engine's operation to generate controllable clock jitter.

7. The anti-side-channel attack encryption computation method integrated into a DPU according to claim 1, characterized in that, The adaptive adjustment instructions generated based on the defense effectiveness evaluation results include: Perform time-domain or frequency-domain feature analysis on the side-channel observation data to obtain the side-channel leakage index for the current encryption operation cycle; The side-channel leakage index is compared with a set of historical evaluation data, which includes a standard power consumption template of the encryption engine collected under a preset security environment, and task execution performance indicators of encryption operations under different perturbation modes. By comparing the deviation between the current leakage index and the standard power consumption template, the defense effectiveness evaluation result is obtained. The defense effectiveness evaluation result indicates whether the current defense is effective or the defense needs to be strengthened. When the defense effectiveness evaluation results indicate that the defense needs to be enhanced, an adaptive adjustment command is generated to increase the amplitude of the simulated disturbance signal or change its frequency.

8. The anti-side-channel attack encryption computation method integrated into a DPU according to claim 7, characterized in that, The comparison of the side-channel leakage index with a set of historical evaluation data is performed by a lightweight reinforcement learning model deployed in the DPU security isolation zone. The lightweight reinforcement learning model receives several sets of historical physical response features, dynamic defense parameters and corresponding side-channel leakage indices as training data, and its output is an optimization suggestion for the parameters of the programmable perturbation command. The optimization suggestion constitutes the core part of the adaptive adjustment command.

9. The anti-side-channel attack encryption computation method integrated into a DPU according to claim 1, characterized in that, The method further includes: Before generating the programmable perturbation instruction, security level configuration information is received from the upper-layer application; Based on the security level configuration information, an initial perturbation mode template is selected from a set of stored perturbation mode templates; The programmable disturbance instruction is generated based on the initial disturbance mode template and the adaptive adjustment instruction.

10. A side-channel attack-resistant encryption computing system integrated into a DPU, characterized in that, The system is used for a side-channel attack prevention encryption computation method integrated into a DPU as described in any one of claims 1-9, the system comprising: The instruction generation module is used to generate a programmable perturbation instruction in response to the encryption operation start instruction, and send the programmable perturbation instruction to a programmable perturbation generator with signal generation function; The physical perturbation and acquisition module is used to control the programmable perturbation generator to inject simulated perturbation signals into the internal physical environment of the DPU chip according to the programmable perturbation instructions, and simultaneously acquire the environmental feedback signals triggered by the simulated perturbation signals to obtain dynamic feedback data. The feature extraction and conversion module is used to perform hardware feature extraction on the dynamic feedback data to obtain several physical response feature quantities, and convert the physical response feature quantities into dynamic defense parameters corresponding to this encryption operation according to the mapping rules used to realize numerical transformation. The encryption execution and observation module is used to control the encryption engine to load the dynamic defense parameters to perform encryption operations, and to synchronously collect side-channel observation data during the encryption operation process. The defense assessment and feedback storage module is used to derive the defense effectiveness assessment result based on the side channel observation data, generate an adaptive adjustment instruction based on the defense effectiveness assessment result, feed the adaptive adjustment instruction back to the generation process of the programmable perturbation instruction, adjust the injection characteristics of the simulated perturbation signal in the subsequent encryption operation cycle, and simultaneously store the physical response characteristics, dynamic defense parameters and side channel observation data of this encryption operation into a set of historical assessment data to realize the continuous updating of the defense knowledge base.

Images