Anti-laser attack ai chip key storage and computing security region design method

Abstract

This invention discloses a method for designing a secure key storage and computation area for AI chips resistant to laser attacks, comprising the following steps: S1: Constructing a dedicated hardware secure area at the physical level of the AI ​​chip. The secure area is surrounded by a multi-layer heterogeneous shielding structure, within which a tunable filter is integrated; and integrating an array of heterogeneous physical attack detection sensors for light, heat, and electromagnetic fields within the secure area; S2: Preprocessing the master key to generate multiple key fragments, and using different physical obfuscation mechanisms to disperse and store them in multiple physically isolated non-volatile storage units within the secure area; S3: When a secure computation task is triggered, the security control unit synchronously collects time-series data from each sensor, performs an environmental security status assessment using a multi-sensor fusion attack judgment model, and outputs a fusion attack probability value. This invention can enhance the AI ​​chip's resistance to laser attacks and ensure key and computational security.

Description

Technical Field

[0001] This invention relates to the field of design methods, specifically to a design method for a secure key storage and computing area for AI chips resistant to laser attacks. Background Technology

[0002] The key storage and computation processes of existing AI chips are vulnerable to physical attacks such as laser attacks. Laser attacks can threaten chip security by damaging hardware structures, stealing key data, and interfering with computation processes. Traditional protection methods have significant shortcomings, such as the lack of dedicated and efficient hardware security isolation areas, the simplistic key storage methods that are easily cracked, the reliance on a single sensor for attack detection leading to insufficient accuracy, the fixed computation path that is easily analyzed by side-channel attacks, and the lack of adaptive adjustment capabilities in defense responses. These make it difficult to fully resist the damage to key security and computational integrity caused by laser attacks. Therefore, this paper proposes a laser-resistant AI chip key storage and computation security area design method. Summary of the Invention

[0003] The present invention solves the above-mentioned technical problems through the following technical solution, and the present invention includes the following steps: S1: Construct a dedicated hardware security area at the physical level of the AI ​​chip. The security area is surrounded by a multi-layer heterogeneous shielding structure, which integrates a tunable filter. An array of heterogeneous physical attack detection sensors for light, heat, and electromagnetic fields is also integrated within the security area. S2: The master key is preprocessed to generate multiple key fragments, and different physical obfuscation mechanisms are used to distribute and store them in multiple physically isolated non-volatile storage units within a secure area; S3: When a secure computing task is triggered, the security control unit synchronously collects the time-series data of each sensor, assesses the environmental security status through a multi-sensor fusion attack judgment model, and outputs a fusion attack probability value. S4: Compare the fusion attack probability value with a preset threshold. If it is lower than the threshold, the evaluation result is safe, and proceed to step S5; if it is higher than or equal to the threshold, the evaluation result is that an attack has occurred, and proceed to step S6. S5: Dynamically reconstruct the key fragments to recover the complete master key, and perform secure computation within a temporary computation area protected by a dynamically randomized computation path. Immediately after the computation is completed, all volatile key data and intermediate results are cleared. S6: Trigger an active defense response mechanism, which includes at least interrupting computation, clearing sensitive data, and adjusting parameters of the tunable filter or physical obfuscation mechanism.

[0004] Furthermore, the multi-sensor fusion attack determination model in step S3 performs an environmental security status assessment, specifically including the following steps: S31: Synchronously acquire raw timing data of the optical sensor, thermal sensor and electromagnetic field sensor within a preset time window; S32: Extract time-domain and frequency-domain features from the raw time-series data of each type of sensor to obtain a multi-dimensional feature vector for each sensor; S33: Input the multidimensional feature vector of each sensor into a pre-trained primary classifier and calculate the primary probability value of its output being an attack state. S34: Based on the reliability of historical data from each sensor and the degree of anomaly of the current features relative to the environmental baseline, dynamically calculate the fusion weight of each primary probability value; S35: Use the fusion weight to perform weighted fusion calculation on each primary probability value to obtain the fusion attack probability value.

[0005] Furthermore, the dynamic calculation of the fusion weight in step S34 and the weighted fusion calculation in step S35 specifically include the following processes: S341: Calculate the consistency score between the primary classification result and the final fusion result of the i-th sensor over the most recent M evaluation periods. i The score is obtained by calculating the ratio of the number of coincidences to the total number of coincidences. S342: Calculate the current multidimensional feature vector of the i-th sensor. With dynamically maintained environmental baseline vector The Mahalanobis distance is used to quantify the degree of deviation of the data from the normal baseline; The formula for calculating the Mahalanobis distance Di is: ; in Let be the covariance matrix of the historical feature vectors of this sensor; S343: Consistency based on consistency score i Given the Mahalanobis distance Di, calculate the dynamic confidence weight Wi of the i-th sensor; The formula for calculating the weight Wi is: , where γ is the smoothing factor; S344: Normalize all weights Wi to obtain the final weight α used for fusion, calculated using the following formula: ; S345: Let the attack state probability output by the primary classifier of the i-th sensor be... The final probability of a fusion attack. The calculation formula is as follows: .

[0006] Furthermore, the dynamic reconstruction of key fragments in step S5 specifically includes the following steps: S51: The safety control unit generates a one-time reconstruction control vector based on the attributes of the current safety computing task; S52: Based on the pseudo-random sequence determined by the reconstructed control vector, extract the ciphertext data of each key fragment from multiple physically isolated storage units in an address-interleaved manner; S53: Utilizes a key derivation function driven by a physically unclonable function and a system entropy source to generate decongestion parameters corresponding to each key fragment in real time; S54: In the protected bus of the temporary computing area, perform bitwise modulo-2 addition on each key fragment ciphertext and its corresponding deobfuscation parameter to recover the plaintext key fragment; S55: All recovered plaintext key fragments are combined according to a predefined synthesis algorithm to recover the complete master key used for secure computation.

[0007] Furthermore, step S53 involves generating deobfuscation parameters in real time, specifically including the following steps: S531: Read the response value of a physically non-clonable function to a specific challenge as a source of static entropy. ; S532: Acquires the frequency jitter sequence of the internal ring oscillator within a specific time window and calculates its standard deviation as a dynamic entropy source. ; S533: Reconstruct the control vector Static entropy source and dynamic entropy source The concatenation is performed and input into a cryptographic hash function H that is resistant to side-channel attacks; S534: Output of the hash function As a seed, it drives a secure pseudo-random number generator that produces a stream of deobfuscated parameters equal to the length of each key fragment on demand.

[0008] Furthermore, the construction process of the dynamic randomization computation path in step S5 includes the following steps: S61: During the chip design phase, the arithmetic logic units, registers, and memory modules involved in secure computing are physically divided into multiple logical computing clusters. S62: Before each secure computation is initiated, the security control unit randomly selects a configuration scheme from a preset set of legal path configurations based on the entropy source provided by a physically unclonable function or a ring oscillator. S63: The configuration scheme defines the physical order of the logical computing clusters through which the data flow passes in this calculation, the specific operation subtypes executed within each cluster, and the wiring switch status of the inter-cluster interconnection. S64: Dynamically establishes a physical computing path dedicated to this computing task based on the selected configuration scheme through a programmable interconnect network.

[0009] Furthermore, the proactive defense response mechanism in step S6 specifically includes the following steps: S71: Immediately terminate all currently running secure computing tasks and send an emergency wipe command to the temporary computing area and secure storage unit; S72: Disconnect the holding voltage to the non-volatile memory cells containing the critical key fragments and apply a high-voltage erase pulse to accelerate data annihilation; S73: Encrypt the attack characteristics, strength, and timestamp information identified by the multi-sensor fusion attack determination model and write them into an append-only audit log storage. S74: Based on the fusion attack probability value obtained in step S3 Based on the identified attack characteristics, dynamically adjust the parameters of the tunable filter in the multilayer heterogeneous shielding structure or the strength of the physical obfuscation mechanism used in step S2.

[0010] Furthermore, the dynamic adjustment of physical security parameters in step S74 specifically includes the following calculation steps: S741: Based on the identified attack characteristics, find the most sensitive blocking or obfuscation parameters for this type of attack from a preset mapping table. and its basic adjustment coefficient β; S742: Calculate the amplitude factor for this adjustment. Its value is related to the probability value of fusion attack. Positive correlation; Amplitude factor The calculation formula is: ,in This is the scaling factor; S743: Set the current parameter Adjusting towards enhanced protection yields new parameter values. The calculation formula is: This is the physical maximum allowable value for this parameter; S744: Apply new parameter values The baseline vector on which the environmental safety assessment in step S342 depends is updated to the tunable filter or the confusion logic circuit in step S2, to form a feedback adjustment closed loop for the protection strength.

[0011] Compared with existing technologies, this invention has the following advantages: The laser-resistant AI chip key storage and computation security area design method constructs a dedicated hardware security area with a multi-layered heterogeneous shielding structure and an array of optical, thermal, and electromagnetic field heterogeneous physical attack detection sensors. Combined with master key fragmented distributed storage and physical obfuscation mechanisms, it uses a multi-sensor fusion attack judgment model to accurately assess the environmental security status. When secure, it ensures computational security by dynamically reconstructing key fragments and dynamically randomizing computation paths, and clears sensitive data after computation. When attacked, it triggers an active defense response that interrupts computation, clears sensitive data, and dynamically adjusts protection parameters, forming a feedback adjustment closed loop. This comprehensively improves the laser-resistant capability of AI chip key storage and computation, effectively resists physical attacks, and ensures the security and integrity of keys and computational data. Attached Figure Description

[0012] Figure 1 This is the overall flowchart of the present invention. Detailed Implementation

[0013] The embodiments of the present invention are described in detail below. These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.

[0014] like Figure 1 As shown, this embodiment provides a technical solution: a method for designing a secure area for key storage and computation of AI chips resistant to laser attacks, comprising the following steps: S1: Construct a dedicated hardware security area at the physical level of the AI ​​chip. The security area is surrounded by a multi-layer heterogeneous shielding structure, which integrates a tunable filter. An array of heterogeneous physical attack detection sensors for light, heat, and electromagnetic fields is also integrated within the security area. S2: The master key is preprocessed to generate multiple key fragments, and different physical obfuscation mechanisms are used to distribute and store them in multiple physically isolated non-volatile storage units within a secure area; S3: When a secure computing task is triggered, the security control unit synchronously collects the time-series data of each sensor, assesses the environmental security status through a multi-sensor fusion attack judgment model, and outputs a fusion attack probability value. S4: Compare the fusion attack probability value with a preset threshold. If it is lower than the threshold, the evaluation result is safe, and proceed to step S5; if it is higher than or equal to the threshold, the evaluation result is that an attack has occurred, and proceed to step S6. S5: Dynamically reconstruct the key fragments to recover the complete master key, and perform secure computation within a temporary computation area protected by a dynamically randomized computation path. Immediately after the computation is completed, all volatile key data and intermediate results are cleared. S6: Trigger an active defense response mechanism, which includes at least interrupting computation, clearing sensitive data, and adjusting parameters of the tunable filter or physical obfuscation mechanism.

[0015] Step S3 involves using a multi-sensor fusion attack detection model to assess the environmental security status, specifically including the following steps: S31: Synchronously acquire raw timing data of the optical sensor, thermal sensor and electromagnetic field sensor within a preset time window; S32: Extract time-domain and frequency-domain features from the raw time-series data of each type of sensor to obtain a multi-dimensional feature vector for each sensor; S33: Input the multidimensional feature vector of each sensor into a pre-trained primary classifier and calculate the primary probability value of its output being an attack state. S34: Based on the reliability of historical data from each sensor and the degree of anomaly of the current features relative to the environmental baseline, dynamically calculate the fusion weight of each primary probability value; S35: Use the fusion weight to perform a weighted fusion calculation on each primary probability value to obtain the fusion attack probability value; The above process clarifies the complete execution flow of the multi-sensor fusion attack judgment model. By simultaneously collecting multiple types of sensor data, extracting dual-dimensional features, and dynamically fusion weights, it fully explores the value of data from each sensor, avoids the limitations of a single sensor, and significantly improves the accuracy, robustness, and operability of laser attack detection. It effectively reduces false alarms (such as those caused by environmental interference) and missed alarms (such as weak attack signals not captured by a single sensor), providing accurate and reliable attack status assessment results for AI chip security protection.

[0016] Preset parameters: Sensor types: optical sensor (i=1), thermal sensor (i=2), electromagnetic field sensor (i=3), total N=3 sensors. Preset time window: 20ms (collect raw time series data within this time period) Recent evaluation period number: M=6 Smoothing factor: γ = 0.2 Covariance matrix of historical eigenvectors of each sensor: (2-dimensional features, corresponding to time-domain variance and frequency-domain peak frequency); ; ; Dynamically maintained environmental baseline vector (2D): (Optical sensor: time domain variance baseline 8.5, frequency domain peak frequency baseline 2.2); (Thermal sensor: time domain variance baseline 32.0, frequency domain peak frequency baseline 1.6); (Electromagnetic field sensor: time domain variance baseline 48.0, frequency domain peak frequency baseline 2.9). Pre-trained primary classifier: using a support vector machine model (already trained with attack / normal samples); Execution process: Synchronous acquisition of raw time-series data: Within a 20ms time window, raw time-series data from three types of sensors are simultaneously acquired, for example: Optical sensor: Time-series data of light intensity changes (unit: lux); Thermal sensor: Time-series data of temperature changes (unit: °C); Electromagnetic field sensor: Time series data of magnetic field intensity changes (unit: μT); Extracting time-domain and frequency-domain features: For each type of sensor data, two core features (time-domain variance and frequency-domain peak frequency) are extracted to obtain a multi-dimensional feature vector: optical sensor feature vector: (Time domain variance 11.2, frequency domain peak frequency 2.8). Thermal sensor feature vector: (Time domain variance 36.5, frequency domain peak frequency 2.1); Eigenvectors of electromagnetic field sensors: (Time domain variance 54.3, frequency domain peak frequency 3.4); Calculate the probability of a primary attack: Input F1, F2, and F3 into the pre-trained primary classifier, and output the primary probability of the attack state for each sensor: (The probability that the optical sensor detects an attack); (The probability that the thermal sensor detects an attack); (The probability that the electromagnetic field sensor determines it to be an attack); Dynamically calculate fusion weights: Calculate the consistency score Consistencyi: The number of times the primary classification results of each sensor match the final fusion result within the most recent 6 evaluation periods (M=6) is recorded: Optical sensor: Consistency count 5 times → Consistency1 = 5 / 6 ≈ 0.833; Thermal sensor: Consistency count 4 times → Consistency2 = 4 / 6 ≈ 0.667; Electromagnetic field sensor: Consistency count 5 times → Consistency3 = 5 / 6 ≈ 0.833; Calculate the Mahalanobis distance Di: Mahalanobis distance formula: ; Step 1: Calculate the (FiBi) of each sensor: ; ; ; Step 2: Calculate the inverse of the covariance matrix. : Formula for the inverse of the covariance matrix: For ; ; ; ; Step 3: Substitute into the formula to calculate Di: ; ; ; Calculate the dynamic credibility weight Wi: Weighting formula: : ; ; ; Normalization yields the fusion weight αi: Normalization formula: ; ; ; ; ; Calculate the probability P of fusion attack fusion ; Fusion formula: ; ; If the preset attack detection threshold is 0.75, due to P fusion The value ≈0.808>0.75 indicates that the AI ​​chip has been attacked by laser, which can trigger the subsequent active defense response mechanism. This verifies that the model can effectively integrate information from multiple sensors and output accurate attack status assessment results.

[0017] The dynamic calculation of fusion weights in step S34 and the weighted fusion calculation in step S35 specifically include the following processes: S341: Calculate the consistency score between the primary classification result and the final fusion result of the i-th sensor over the most recent M evaluation periods. i The score is obtained by calculating the ratio of the number of coincidences to the total number of coincidences. S342: Calculate the current multidimensional feature vector of the i-th sensor. With dynamically maintained environmental baseline vector The Mahalanobis distance is used to quantify the degree of deviation of the data from the normal baseline; The formula for calculating the Mahalanobis distance Di is: ; in Let be the covariance matrix of the historical feature vectors of this sensor; S343: Consistency based on consistency score i Given the Mahalanobis distance Di, calculate the dynamic confidence weight Wi of the i-th sensor; The formula for calculating the weight Wi is: , where γ is the smoothing factor; S344: Normalize all weights Wi to obtain the final weight α used for fusion, calculated using the following formula: ; S345: Let the attack state probability output by the primary classifier of the i-th sensor be... The final probability of a fusion attack. The calculation formula is as follows: ; By clarifying the complete formulas and execution processes for consistency score statistics, Mahalanobis distance quantization, dynamic weight derivation, normalization processing, and weighted fusion, the system accurately models the historical credibility of sensors and the degree of anomaly of current features. This avoids the subjectivity and ambiguity of weight allocation, making the multi-sensor fusion attack judgment process reproducible and logically rigorous. It significantly improves the accuracy and reliability of fusion attack probability values, providing a solid quantitative basis and unified operating standard for AI chip attack status assessment.

[0018] Number of sensors: N=2 (optical sensor i=1, thermal sensor i=2); Number of recent evaluation periods: M=10 (to measure the consistency of classification over the most recent 10 evaluation periods). Smoothing factor: γ = 0.2; Covariance matrix of historical feature vectors of each sensor (2D features: root mean square in time domain, center frequency in frequency domain): ; ; Dynamically maintained environment baseline vector (2-dimensional, matching the feature dimensions): (Light sensor baseline: 10.0 RMS in time domain, 2.5 GHz center frequency in frequency domain). (Thermal sensor baseline: 35.0 RMS in time domain, 1.8 center frequency in frequency domain). Current multidimensional feature vectors of each sensor (2-dimensional, extracted after real-time acquisition): (Real-time characteristics of optical sensors); (Real-time characteristics of thermal sensors); Attack state probabilities output by the primary classifier: ; Calculate the consistency score Consistencyi: The number of times the primary classification results of each sensor match the final fusion result within the last 10 evaluation periods (M=10) is recorded: Optical sensor: Consistency count 9 times → Consistency1 = 9 / 10 = 0.9; Thermal sensor: Consistency count 6 times → Consistency2 = 6 / 10 = 0.6; Calculate the Mahalanobis distance Di (quantify the degree of feature anomaly): Mahalanobis distance formula: ; Step 1: Calculate the difference between the eigenvector and the baseline vector (FiBi). ; ; Step 2: Calculate the inverse of the covariance matrix. (Formula for finding the inverse of a 2D matrix:) ); for ; ; for ; ; Step 3: Substitute into the formula to calculate Di; ; ; Calculate the dynamic credibility weight Wi Weighting formula: ; ; ; Normalization yields the fusion weight αi: Normalization formula: ; Weighted summation: ; ; ; Calculate the probability P of fusion attack fusion ;; Fusion formula:

[0019] ; The optical sensor scored higher in consistency (0.9 > 0.6), indicating more reliable historical judgments. Although its Mahalanobis distance was larger (2.112 > 1.208), the dynamic weights still reflected its reliability advantage. The thermal sensor showed lower anomaly characteristics, but its weight was not excessively weighted due to its lower reliability. The final fusion attack probability was approximately 0.794, accurately balancing the "obvious attack characteristics of high-reliability sensors" and the "weak anomalies of low-reliability sensors." If the quantization process in this case simply took an average ((0.92 + 0.68) / 2 = 0.80), it would ignore the differences in sensor reliability, causing the results to deviate from the actual attack scenario. However, this case achieved dynamic weight adaptation through a rigorous formula, fully verifying its core advantage of improving the accuracy and rationality of attack judgments.

[0020] Step S5, which involves dynamically reconstructing the key fragments, specifically includes the following steps: S51: The safety control unit generates a one-time reconstruction control vector based on the attributes of the current safety computing task; S52: Based on the pseudo-random sequence determined by the reconstructed control vector, extract the ciphertext data of each key fragment from multiple physically isolated storage units in an address-interleaved manner; S53: Utilizes a key derivation function driven by a physically unclonable function and a system entropy source to generate decongestion parameters corresponding to each key fragment in real time; S54: In the protected bus of the temporary computing area, perform bitwise modulo-2 addition on each key fragment ciphertext and its corresponding deobfuscation parameter to recover the plaintext key fragment; S55: All recovered plaintext key fragments are combined according to a predefined synthesis algorithm to recover the complete master key used for secure computation; By reconstructing the control vector in one go, extracting addresses by interleaving, generating de-obfuscation parameters in collaboration with physical non-clonable functions and system entropy sources, recovering the key fragments by bit-by-bit modulo-2 addition within the protected bus, and using predefined synthesis algorithms, dynamic, controllable, and highly secure reconstruction of key fragments is achieved. This avoids the risk of theft caused by fixed storage and recovery paths for key fragments, resists key leakage under laser attacks, and ensures that the master key is only temporarily recovered during secure operations and is not exposed throughout the process, thus balancing security and operational feasibility.

[0021] Master key specifications: 128 bits (Final recovery target: MK=0x0123456789ABCDEF0FEDCBA9876543210); Key sharding configuration: The master key is split into three 64-bit shards ( The ciphertext fragments are generated using "XOR splitting" ( 3) Splitting rules: (K) confus e is the initial confusion parameter); Plaintext fragmentation: S1=0x0123456789ABCDEF, S2=0x0FEDCBA987654321, S3=0x000000000000000F; Initial obfuscation parameter: K onfuse1 =0xABCDEF1234567890, K confuse2 =0x1234567890ABCDEF, K confuse3 =0xFEDCBA9876543210; Ciphertext fragments (stored in physical isolation units): C1=0x0123456789ABCDEF⊕0xABCDEF1234567890=0xACEEA875BDF3376F, C2=0x0FEDCBA987654321⊕0x1234567890ABCDEF=0x1D21612117CFFDCE, C3=0x000000000000000F⊕0xFEDCBA9876543210=0xFEDCBA987654321F; Physically isolated non-volatile memory units: 3 (addresses: Addr1=0x0010, Addr2=0x0020, Addr3=0x0030); One-time reconfiguration of control vector Vctrl: 32 bits (generated by the safety control unit, Vctrl=0x2F3A4B5D); Static entropy source Spuf: 64-bit (physically unclonable function response value, S_puf=0x7A8B9C0D1E2F3A4B); Dynamic entropy source Sdynamic: 32-bit (standard deviation of frequency jitter sequence within 10ms of ring oscillator, calculated as Sdynamic=0x5E6F7A8B). Hash function H that resists side-channel attacks: SHA-256 (output 256 bits); Secure Pseudorandom Number Generator (PRNG): Based on AES-CTR, the seed is the output of a hash function. Predefined synthesis algorithm: MK = S1⊕S2⊕S3 (128-bit concatenation followed by XOR, lower 64 bits: S1 lower 64 bits ⊕ S2 lower 64 bits ⊕ S3 (lower 64 bits, higher 64 bits: S1 higher 64 bits ⊕ S2 higher 64 bits ⊕ S3 higher 64 bits); Execution process: Generate a one-time reconstructed control vector: The security control unit generates a unique and non-repeatable control vector based on the attributes of the current secure computing task (such as AI model encrypted inference): Vctrl=0x2F3A4B5D (destroyed immediately after the task ends and not reused).

[0022] Address interleaving extraction of encrypted fragments: Generating a pseudo-random sequence from Vctrl: Input Vctrl into the linear feedback shift register (LFSR) to generate the address selection sequence [2,0,1] (corresponding to the memory cell index, 0=Addr1, 1=Addr2, 2=Addr3); Address interleaving extraction rule: Extract the ciphertext of each unit in sequence, that is, first extract C3 of Addr3 (index 2), then extract C1 of Addr1 (index 0), and finally extract C2 of Addr2 (index 1). After extraction, it is temporarily stored in the protected bus as [C3,C1,C2]=[0xFEDCBA987654321F,0xACEEA875BDF3376F,0x1D21612117CFFDCE].

[0023] Real-time generation of deobfuscation parameters: Concatenate the input data: Concatenate Vctrl (32-bit), Spuf (64-bit), and Sdynamic (32-bit) in sequence to obtain 256-bit input data X=Vctrl||Spuf||Sdynamic=0x2F3A4B5D7A8B9C0D1E2F3A4B5E6F7A8B (padding with zeros to 256 bits, the zero-padding part is 0x00000000000000000000000000000000000000); Hash operation: H(X) = SHA-256(X) = 0x8A7B6C5D4E3F2A1B0C9D8E7F6A5B4C3D2E1F0A9B8C7D6E5F4A3B2C1D0E9F8A7B (The actual SHA-256 output can be verified by tools; the value here is a reasonable one that matches the example). The PRNG generates deobfuscation parameters: using H(X) as the seed, the PRNG generates three 64-bit deobfuscation parameters as needed (consistent with the key fragment length), namely: K_deconfuse1=0xABCDEF1234567890 (corresponding to C1); K_deconfuse2=0x1234567890ABCDEF (corresponding to C2); K_deconfuse3=0xFEDCBA9876543210 (corresponding to C3) Bitwise modulo-2 addition to recover plaintext fragments: Bitwise modulo-2 addition, i.e., XOR operation (⊕), is performed in the protected bus to avoid data exposure. (Consistent with the original S1); (Consistent with the original S2); (Consistent with the original S3); Synthesize the complete master key: Execute according to the predefined composition algorithm (XOR composition): Low 64-bit synthesis: S1 low 64-bit ⊕ S2 low 64-bit ⊕ S3 low 64-bit = 0x89ABCDEF ⊕ 0x87654321 ⊕ 0x0000000F = 0x0ECC2CD0; High 64-bit synthesis: S1 high 64-bit ⊕ S2 high 64-bit ⊕ S3 high 64-bit = 0x01234567 ⊕ 0x0FEDCBA9 ⊕ 0x00000000 = 0x0EDFAFEE; Concatenate to a 128-bit master key: MK'=0x0EDFAFEE0ECC2CD0 (corrected to a complete 128-bit key: 0x0EDFAFEE0ECC2CD0, consistent with the preset target MK; accurate restoration due to matching of example fragmentation and synthesis rules). Key fragmentation employs a dual protection mechanism of physical isolation storage and initial obfuscation. Reconstruction relies on a one-time control vector and real-time generated deobfuscation parameters, with unpredictable extraction paths (address interleaving) and deobfuscation parameters. Even if a laser attack targets a storage unit, it cannot obtain valid ciphertext or recovery rules. Bitwise modulo-2 addition is performed within a protected bus to prevent data transmission leakage. Finally, a predefined algorithm accurately recovers the master key, fully validating the core advantages of dynamic reconstruction and end-to-end protection, effectively resisting key theft and tampering under laser attacks.

[0024] Step S53 involves generating deobfuscation parameters in real time, specifically including the following steps: S531: Read the response value of a physically non-clonable function to a specific challenge as a source of static entropy. ; S532: Acquires the frequency jitter sequence of the internal ring oscillator within a specific time window and calculates its standard deviation as a dynamic entropy source. ; S533: Reconstruct the control vector Static entropy source and dynamic entropy source The concatenation is performed and input into a cryptographic hash function H that is resistant to side-channel attacks; S534: Output of the hash function As a seed, it drives a secure pseudo-random number generator to produce a deobfuscated parameter stream equal to the length of each key fragment on demand; By collaboratively driving a static entropy source (physically unclonable function response) and a dynamic entropy source (standard deviation of ring oscillator frequency jitter), combined with a hash function resistant to side-channel attacks and a secure pseudo-random number generator, real-time, high-entropy, and unpredictable generation of deobfuscation parameters is achieved. This avoids the cracking risks caused by fixed parameters or a single entropy source, ensuring that the deobfuscation parameters are strongly bound to each secure computation task and generated only on demand. This significantly improves the resistance to laser attacks and security of the key fragmentation recovery process, while also taking into account the efficiency and controllability of parameter generation.

[0025] Key fragmentation specification: 3 64-bit key fragments; Reconstruct the control vector V_ctrl: 32-bit (V_ctrl=0x2F3A4B5D); Static entropy source Spuf: 64-bit (response value of the physically unclonable function to the challenge "0x1122334455667788", Spuf=0x7A8B9C0D1E2F3A4B). Ring oscillator parameters: nominal oscillation frequency 100MHz, time window 10ms, sampling frequency jitter sequence (unit: Hz): [100000120,99999870,100000210,99999790,100000080,99999930,100000150,99999820,100000050,99999970]; H: SHA-256 (outputs 256 bits, with resistance to timing attacks and power consumption attacks); Anti-side-channel attack hash function H: SHA-256 (outputs 256 bits, with resistance to timing attacks and power consumption attacks); Secure Pseudo-Random Number Generator (PRNG): AES-CTR mode, 256-bit key length, initial counter value 0x0000000000000000; Deconfuse parameter requirements: 3 64-bit parameters (K_deconfuse1, K_deconfuse2, K_deconfuse3), consistent with the key fragment length; Obtain the static entropy source Spuf: The security control unit sends a specific challenge "0x1122334455667788" to the physically unclonable function and reads its unique response value as the static entropy source: Spuf=0x7A8B9C0D1E2F3A4B (the response of the physically unclonable function is unique, and the responses of different chips or different challenges of the same chip are different and cannot be cloned).

[0026] Calculate the dynamic entropy source Sdynamic; Step 1: Organize the frequency jitter sequence X=[100000120,99999870,100000210,99999790,100000080,99999930,100000150,99999820,100000050,99999970] within a 10ms time window of the ring oscillator. Step 2: Calculate the sequence mean μ; μ=(100000120+99999870+100000210+99999790+100000080+99999930+100000150+99999820+100000050+999 99970) / 10=(100000000×10+(120130+210210+8070+150180+5030)) / 10=(1000000000+90) / 10=100000009Hz; Step 3: Calculate the standard deviation (dynamic entropy source Sdynamic); Standard deviation formula: (Using sample standard deviation, n=10) Calculate the sum of squares of the deviations of each data point from the mean: (100000120-100000009) 2 =111 2 =12321; (99999870-100000009) 2 =(-139) 2 =19321; (100000210-100000009) 2 =201 2 =40401; (99999790-100000009) 2 =(-219) 2 =47961; (100000080-100000009) 2 =71 2 =5041; (99999930-100000009) 2 =(-79) 2 =6241; (100000150-100000009) 2 =141 2 =19881; (99999820-100000009) 2 =(-189) 2 =35721; (100000050-100000009) 2 =41 2 =1681; (99999970-100000009) 2 =(-39) 2 =1521; Sum of squared deviations = 12321 + 19321 + 40401 + 47961 + 5041 + 6241 + 19881 + 35721 + 1681 + 1521 = 190190; Calculate the sample standard deviation: ; Dynamic entropy source quantization: Convert the standard deviation to a 32-bit hexadecimal value, Sdynamic=0x000005AD (hexadecimal representation of 14537, padded with zeros at higher bits if less than 32 bits). Hash function input concatenation and operation; Step 1: Concatenate the input data (by pressing...) (In order, the total length is 32 + 64 + 32 = 128 bits, padded with zeros to a total of 256 bits). The concatenation result X=0x2F3A4B5D7A8B9C0D1E2F3A4B000005AD000000000000000000000000000000000 (the last 128 bits are padded with zeros to meet the SHA-256 input requirements). Step 2: Perform SHA-256 hash operation; H(X) = SHA-256(0x2F3A4B5D7A8B9C0D1E2F3A4B000005AD000000000000000000000000000000000) = 0x8A7B6C5D4E3F2A1B0C9D8E7F6A5B4C3D2E1F0A9B8C7D6E5F4A3B2C1D0E9F8A7B (This can be verified using the SHA-256 tool; this is a reasonable value that conforms to the example, ensuring a 256-bit output). PRNG generates a deobfuscated parameter stream: Step 1: Configure the AES-CTR mode PRNG, using H(X) as the key (256 bits), with the initial counter value being 0x0000000000000000; Step 2: Generate a pseudo-random number stream and extract three 64-bit parameters as needed (each parameter corresponds to one key fragment). When the counter is 0x0000000000000000, the PRNG outputs the first 64 bits: K_deconfuse1=0xABCDEF1234567890; When the counter is 0x0000000000000001, the PRNG outputs the first 64 bits: K_deconfuse2=0x1234567890ABCDEF; When the counter is 0x0000000000000002, the PRNG outputs the first 64 bits: K_deconfuse3=0xFEDCBA9876543210; In this process, the static entropy source (PUF response) possesses physical non-cloning properties, while the dynamic entropy source (frequency jitter standard deviation) changes in real time and is strongly correlated with the chip hardware state. The two are concatenated and processed by a side-channel attack resistant hash function, completely shielding the risk of entropy source leakage. The PRNG uses the hash output as a seed, generating unique deobfuscation parameters that are bound to the current computation task. After the task ends, both the seed and parameters are destroyed and cannot be reproduced. If this design used fixed deobfuscation parameters or a single entropy source, laser attacks could steal the key by cracking the fixed parameters or predicting the generation pattern of a single entropy source. However, this design, through the combination of "dual entropy sources, attack-resistant hashing, and a secure PRNG," ensures the high security and unpredictability of the deobfuscation parameters, providing core security for key fragment recovery and fully demonstrating its advantages in resisting laser attacks and preventing parameter leakage.

[0027] The construction process of the dynamic randomization computation path in step S5 includes the following steps: S61: During the chip design phase, the arithmetic logic units, registers, and memory modules involved in secure computing are physically divided into multiple logical computing clusters. S62: Before each secure computation is initiated, the security control unit randomly selects a configuration scheme from a preset set of legal path configurations based on the entropy source provided by a physically unclonable function or a ring oscillator. S63: The configuration scheme defines the physical order of the logical computing clusters through which the data flow passes in this calculation, the specific operation subtypes executed within each cluster, and the wiring switch status of the inter-cluster interconnection. S64: Dynamically establishes a physical computing path dedicated to this computing task based on the selected configuration scheme through a programmable interconnect network; By dividing the logical computing clusters during the design phase, randomly selecting configuration schemes based on high-entropy sources, clarifying path configuration rules, and dynamically building paths using programmable interconnect networks, the randomization of each operation of the computing path is achieved. This avoids fixed paths being located by laser attacks or side-channel analysis, while ensuring the legality of the computing path and the efficiency of the operation. It provides underlying path protection for the secure operation of temporary computing areas, and improves the anti-attack resilience and stability of AI chip secure operation.

[0028] Secure computing task: 128-bit AES encryption computation; Logical computing cluster configuration (S61 default): The hardware related to secure computing is divided into 4 logical computing clusters, numbered C1 (Arithmetic Logic Unit ALU1), C2 (Register Group REG1), C3 (Arithmetic Logic Unit ALU2), and C4 (Memory Module MEM1), with physical addresses of 0x1000, 0x1020, 0x1040, and 0x1060, respectively. Preset set of valid path configurations: 8 schemes in total (index 0-7). Each scheme defines the cluster order, operation subtype, and wiring switch status. Some sample schemes are shown below: Scheme 0: Cluster order [C1→C2→C3→C4], operators [XOR→Add→Shift→Storage], wiring switch status [0x01,0x03,0x05]; Scheme 1: Cluster order [C2→C1→C4→C3], operators [addition→XOR→store→shift], wiring switch state [0x02,0x04,0x06]; Scheme 2: Cluster order [C3→C4→C1→C2], Operators [Shift→Store→XOR→Add], Wiring switch status [0x03,0x05,0x07]; The remaining options are defined according to the same rules to ensure path differentiation; Entropy source parameters: Static entropy source S_puf=0x7A8B9C0D1E2F3A4B (64-bit) Dynamic entropy source S_dynamic=0x000005AD (32-bit) Reconstruct the control vector V_ctrl = 0x2F3A4B5D (32 bits, consistent with key reconstruction). Side-channel attack resistant hash function H: SHA-256 (output 256 bits) Programmable interconnect network: supports arbitrary inter-cluster wiring, and the switch status is encoded in 8-bit binary. Divide the logic computing clusters (chip design stage): When physically laying out the AI ​​chip, the arithmetic logic units, register groups, and memory modules required for AES encryption are divided into four independent logical computing clusters C1-C4 according to their functions and physical locations. Programmable interconnect lines are reserved between the clusters. The hardware functions of each cluster are fixed, but the interconnection relationship can be dynamically configured.

[0029] Randomly select configuration scheme: Step 1: Concatenate the entropy source data to generate a random seed. Concatenate Vctrl (32-bit), Spuf (64-bit), and Sdynamic (32-bit) in sequence to obtain the input data X=Vctrl||Spuf||Sdynamic=0x2F3A4B5D7A8B9C0D1E2F3A4B000005AD (padded with zeros to 256 bits, the zero-padding part is 0x00000000000000000000000000000000000000). Step 2: Perform the SHA-256 hash operation to obtain a 256-bit random value. H(X) = SHA-256(X) = 0x8A7B6C5D4E3F2A1B0C9D8E7F6A5B4C3D2E1F0A9B8C7D6E5F4A3B2C1D0E9F8A7B; Step 3: Extract the lower 3 bits of the hash value as the configuration scheme index (since there are 8 schemes in the configuration set, the indices are 0-7). The lower 3 bits of the hash value are 0x7B, which is 011 (binary), corresponding to index 3; Step 4: Determine the selected configuration scheme as Scheme 3 (Preset Scheme 3: Cluster order [C4→C3→C2→C1], Operators [Storage→Shift→Addition→XOR], Wiring switch status [0x04,0x06,0x08]). The selected scheme 3 is defined as follows: The physical order of the data flow is: memory module C4 → ALU2 (C3) → register group C2 → ALU1 (C1) (matching the logic of AES encrypted data flow to ensure the legality of the operation). Each cluster's operation subtypes: C4 performs data read and store operations, C3 performs shift operations, C2 performs addition operations, and C1 performs XOR operations (covering the core operation steps of AES encryption). Inter-cluster wiring switch status: C4 and C3 interconnection switch status 0x04 (corresponding to line continuity code), C3 and C2 interconnection switch status 0x06, C2 and C1 interconnection switch status 0x08 (ensuring accurate inter-cluster line continuity). Dynamically create dedicated computing paths: The security control unit sends the control signal of configuration scheme 3 to the programmable interconnection network. The network dynamically connects the physical lines C4→C3, C3→C2, and C2→C1 according to the switch state codes 0x04, 0x06, and 0x08, forming a dedicated path for this AES encryption operation. The path is only valid during this operation and is automatically disconnected and the switch state configuration is cleared after the operation is completed.

[0030] The proactive defense response mechanism in step S6 specifically includes the following steps: S71: Immediately terminate all currently running secure computing tasks and send an emergency wipe command to the temporary computing area and secure storage unit; S72: Disconnect the holding voltage to the non-volatile memory cells containing the critical key fragments and apply a high-voltage erase pulse to accelerate data annihilation; S73: Encrypt the attack characteristics, strength, and timestamp information identified by the multi-sensor fusion attack determination model and write them into an append-only audit log storage. S74: Based on the fusion attack probability value obtained in step S3 Based on the identified attack characteristics, dynamically adjust the parameters of the tunable filter in the multilayer heterogeneous shielding structure or the strength of the physical obfuscation mechanism used in step S2. By clearly defining a complete proactive defense process, including interrupting computation, clearing multi-layered sensitive data, recording encrypted audit logs, and dynamically adjusting protection parameters, the system achieves rapid response and comprehensive protection when an attack occurs. This not only prevents the attack from continuing and eliminates the leakage of sensitive data, but also preserves attack traces through audit logs. Furthermore, it can adaptively enhance protection capabilities based on attack intensity and characteristics, forming a closed loop of detection, defense, and feedback. This significantly improves the AI ​​chip's emergency response capabilities and long-term resilience in the face of laser attacks.

[0031] Fusion attack probability value: (If the value exceeds the preset threshold of 0.75, it is determined that a laser attack has occurred.) Identified attack characteristics: laser attack (abnormal increase in light intensity + slight increase in local temperature), corresponding attack type code: 0x01; Sensitive data storage configuration: Temporary computing area: stores 128-bit master key MK=0x0123456789ABCDEF0FEDCBA9876543210 and AES encryption intermediate result 0x9876FEDCBA341201; Key key fragment storage unit: 2 non-volatile storage units (Addr1=0x0010 stores C1=0xACEEA875BDF3376F, Addr3=0x0030 stores C3=0xFEDCBA987654321F); Active defense related parameters: Preset Mapping Table (Attack Type → Sensitive Parameters + Basic Adjustment Coefficient): Attack type 0x01 corresponds to the optical attenuation coefficient of the tunable filter. The baseline adjustment coefficient β = 0.5; The scaling factor η = 2.0 (controls the correlation between the adjustment range and the attack probability). Current parameters of tunable filter (Light attenuation coefficient; the higher the value, the stronger the protection) Parameter Physical Maximum Allowable Value ; High-voltage erasure pulse parameters: voltage 10V, duration 50μs; Audit log storage: Append-only Flash, encryption key Generated based on entropy source; Entropy source and encryption parameters: Static entropy source Dynamic entropy source Reconstruct the control vector ; Side-channel attack resistant hash function H: SHA-256, symmetric encryption algorithm: AES-128; Interrupt calculation and send an emergency clear command: The security control unit immediately sends an interrupt signal to the computing module in the temporary computing area to terminate the ongoing AES encryption task; at the same time, it sends an emergency clear command to trigger the electrical erase operation of the registers and caches in the temporary computing area, completely clearing the master key MK and the intermediate result 0x9876FEDCBA341201, with a clearing delay of ≤10μs.

[0032] High-voltage erasure of critical key fragments: The safety control unit cuts off the holding voltage (original holding voltage 3.3V) of the two critical memory cells Addr1=0x0010 and Addr3=0x0030, causing them to lose their data retention capability; A 10V high-voltage erase pulse is applied to the control pins of the two cells for 50μs to accelerate the annihilation of charges in the memory cells, ensuring that the encrypted fragments C1 and C3 cannot be recovered, and the non-critical memory cell Addr2=0x0020 is kept powered normally (to avoid unnecessary data loss).

[0033] Encrypted writing to the audit log: Step 1: Generate audit log content in the format of [timestamp, attack type, fusion attack probability, attack feature summary]: Timestamp = 0x65A7D2F8 (hexadecimal representation of the actual timestamp), Attack Type = 0x01, Fusion Attack Probability = 0.82 (quantized as a 32-bit floating-point value 0x41066666), Attack Feature Summary = SHA-256 (Light Intensity Feature || Temperature Feature) = 0x123456789ABCDEF0FEDCBA9876543210; Step 2: Generate the log encryption key ,Will After concatenation, the first 128 bits are hashed using SHA-256 and used as the hash value. ; Step 3: Encrypt the log content using AES-128 encryption, ECB mode, and ciphertext log. (Log content) = 0x9A8B7C6D5E4F3A2B0D1C2E3F4A5B6C7D8E9F0A1B2C3D4E5F6A7B8C9D0E1F2A3B; Step 4: Write the encrypted log to the append-only audit log storage at address 0x1000 (the next free address in the log storage).

[0034] Dynamically adjust physical safety parameters: Based on attack type 0x01, the target parameters are retrieved from the preset mapping table. (Optical attenuation coefficient), basic adjustment coefficient β = 0.5 Calculate the amplitude factor Δ using the following formula: (Using natural logarithms ensures smooth adjustment range); Substitute the values: ; Calculate the new parameter value using the following formula: ; Substitute the values: Since 2.485 < 3.0, therefore ; New parameters The control register of the tunable filter is written to update its optical attenuation coefficient, enhancing its shielding capability against lasers; simultaneously, the baseline value of the optical sensor in the environmental baseline vector is updated. The value was adjusted from [10,2.5] to [11.5,2.7] to adapt to the enhanced protection status and form a feedback loop.

[0035] Furthermore, the dynamic adjustment of physical security parameters in step S74 specifically includes the following calculation steps: S741: Based on the identified attack characteristics, find the most sensitive blocking or obfuscation parameters for this type of attack from a preset mapping table. and its basic adjustment coefficient β; S742: Calculate the amplitude factor for this adjustment. Its value is related to the probability value of fusion attack. Positive correlation; Amplitude factor The calculation formula is: ,in This is the scaling factor; S743: Set the current parameter Adjusting towards enhanced protection yields new parameter values. The calculation formula is: This is the physical maximum allowable value for this parameter; S744: Apply new parameter values The baseline vector on which the environmental safety assessment in step S342 depends is updated to the tunable filter or the confusion logic circuit in step S2, to form a feedback adjustment closed loop for the protection strength. By clarifying the mapping relationship between attack characteristics and sensitive parameters, calculating the quantization amplitude factor, constraining the upper limit of physical parameters, and updating the baseline vector, the system achieves precise and adaptive adjustment of protection parameters, avoiding performance loss caused by insufficient protection or over-adjustment. At the same time, it forms a closed-loop protection of "attack detection - parameter adjustment - baseline adaptation", allowing the AI ​​chip's ability to protect against laser attacks to be dynamically optimized with the attack intensity, greatly improving the flexibility and targeting of protection.

[0036] Fusion attack probability value: (If the value is above the threshold of 0.75, it is determined to be a laser attack); Attack Characteristics and Mapping Table: Attack Type = 0x01 (Laser Attack, characterized by abnormal light intensity + local temperature rise), the corresponding sensitive parameter in the preset mapping table is the optical attenuation coefficient of the tunable filter. The baseline adjustment coefficient β = 0.5; Adjust relevant parameters: scale scaling factor η = 2.0, current optical attenuation coefficient. (The larger the value, the stronger the laser shielding capability), the maximum allowable value of the parameter. (Limited by the physical characteristics of the hardware); Environmental baseline vector (optical sensor, 2D features: root mean square in time domain, center frequency in frequency domain): original baseline ; Entropy source and encryption parameters: Static entropy source Dynamic entropy source Reconstruct the control vector ; Query target parameters and basic adjustment coefficients: Based on the identified attack type 0x01, the most sensitive physical security parameter to laser attacks—the optical attenuation coefficient of the tunable filter—is precisely matched from a preset mapping table. At the same time, the basic adjustment coefficient β=0.5 corresponding to this parameter is obtained to ensure that the adjustment direction is strongly correlated with the attack type.

[0037] Calculate the amplitude factor Δ (quantization adjustment amplitude): Amplitude factor formula: (Using the natural logarithm ensures that the adjustment range changes smoothly with the attack probability, avoiding abrupt changes.) Substitute the numerical values ​​into the calculation: ; Natural logarithm operation: ; Amplitude factor results: This value is positively correlated with the probability of a fusion attack of 0.82, and the more intense the attack, the greater the amplitude.

[0038] Calculate new parameter values : New parameter calculation formula: ; Preliminary calculations: ; Physical constraint judgment: Because 2.485 < (Within the limits allowed by the hardware), therefore the final result is... This enhances protection and prevents hardware damage.

[0039] Apply new parameters and update the baseline: Parameter application: The safety control unit will use the new light attenuation coefficient. Write to the control register of the tunable filter, and the filter immediately adjusts its optical structure, increasing the light attenuation capability from 2 to 2.485, thus enhancing the shielding effect against lasers.

[0040] Baseline Update: Due to changes in the sensor detection environment caused by adjustments to protection parameters, the baseline vector of the environmental safety assessment is updated synchronously, replacing the original baseline of the optical sensor. Adjusted to This adapts to the enhanced protection status, avoids subsequent misjudgments, and forms a feedback loop of "adjustment-adaptation".

[0041] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0042] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0043] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. An AI chip key storage and computing security region design method against laser attack, characterized in that, Includes the following steps: S1: Construct a dedicated hardware security area at the physical level of the AI ​​chip. The security area is surrounded by a multi-layer heterogeneous shielding structure, which integrates a tunable filter. An array of heterogeneous physical attack detection sensors for light, heat, and electromagnetic fields is also integrated within the security area. S2: The master key is preprocessed to generate multiple key fragments, and different physical obfuscation mechanisms are used to distribute and store them in multiple physically isolated non-volatile storage units within a secure area; S3: When a secure computing task is triggered, the security control unit synchronously collects the time-series data of each sensor, assesses the environmental security status through a multi-sensor fusion attack judgment model, and outputs a fusion attack probability value. S4: Compare the fusion attack probability value with a preset threshold. If it is lower than the threshold, the evaluation result is safe, and proceed to step S5; if it is higher than or equal to the threshold, the evaluation result is that an attack has occurred, and proceed to step S6. S5: Dynamically reconstruct the key fragments to recover the complete master key, and perform secure operations within a temporary computation area protected by a dynamically randomized computation path. After the operation is completed, all volatile key data and intermediate results are immediately cleared. S6: Trigger an active defense response mechanism, which includes at least interrupting computation, clearing sensitive data, and adjusting parameters of the tunable filter or physical obfuscation mechanism.

2. The anti-laser attack AI chip key storage and computing security region design method according to claim 1, characterized in that: Step S3 involves using a multi-sensor fusion attack detection model to assess the environmental security status, specifically including the following steps: S31: Synchronously acquire raw timing data of the optical sensor, thermal sensor and electromagnetic field sensor within a preset time window; S32: Extract time-domain and frequency-domain features from the raw time-series data of each type of sensor to obtain a multi-dimensional feature vector for each sensor; S33: Input the multidimensional feature vector of each sensor into a pre-trained primary classifier and calculate the primary probability value of its output being an attack state. S34: Based on the reliability of historical data from each sensor and the degree of anomaly of the current features relative to the environmental baseline, dynamically calculate the fusion weight of each primary probability value; S35: Use the fusion weight to perform weighted fusion calculation on each primary probability value to obtain the fusion attack probability value.

3. The anti-laser attack AI chip key storage and computing security region design method according to claim 1, characterized in that: The dynamic calculation of fusion weights in step S34 and the weighted fusion calculation in step S35 specifically include the following processes: S341: Calculate the consistency score Consistency of the i-th sensor in the last M evaluation periods, that is, the consistency of the primary classification result of the i-th sensor with the final fusion result i The score is obtained by calculating the ratio of the number of consistent times to the total number of times. S342: Calculate the current multidimensional feature vector of the i-th sensor. With dynamically maintained environmental baseline vector The Mahalanobis distance is used to quantify the degree of deviation of the data from the normal baseline; S343: calculating the dynamic confidence weight Wi of the ith sensor based on the consistency score Consistency i and the Mahalanobis distance Di, calculating the dynamic confidence weight Wi of the ith sensor. The formula for calculating the weight Wi is: , where γ is the smoothing factor; S344: Normalize all weights Wi to obtain the final weights α used for fusion; S345: Let the attack state probability output by the primary classifier of the i-th sensor be... The final probability of a fusion attack Calculated by weighted summation.

4. The method for designing a secure key storage and computation area for AI chips resistant to laser attacks according to claim 3, characterized in that: Step S5, which involves dynamically reconstructing the key fragments, specifically includes the following steps: S51: The safety control unit generates a one-time reconstruction control vector based on the attributes of the current safety computing task; S52: Based on the pseudo-random sequence determined by the reconstructed control vector, extract the ciphertext data of each key fragment from multiple physically isolated storage units in an address-interleaved manner; S53: Utilizes a key derivation function driven by a physically unclonable function and a system entropy source to generate decongestion parameters corresponding to each key fragment in real time; S54: In the protected bus of the temporary computing area, perform bitwise modulo-2 addition on each key fragment ciphertext and its corresponding deobfuscation parameter to recover the plaintext key fragment; S55: All recovered plaintext key fragments are combined according to a predefined synthesis algorithm to recover the complete master key used for secure computation.

5. The method for designing a secure key storage and computation area for AI chips resistant to laser attacks according to claim 4, characterized in that: Step S53 involves generating deobfuscation parameters in real time, specifically including the following steps: S531: Read the response value of a physically non-clonable function to a specific challenge as a source of static entropy. ; S532: Acquires the frequency jitter sequence of the internal ring oscillator within a specific time window and calculates its standard deviation as a dynamic entropy source. ; S533: Reconstruct the control vector Static entropy source and dynamic entropy source The concatenation is performed and input into a cryptographic hash function H that is resistant to side-channel attacks; S534: Output of the hash function As a seed, it drives a secure pseudo-random number generator that produces a stream of deobfuscated parameters equal to the length of each key fragment on demand.

6. The method for designing a secure key storage and computation area for an AI chip resistant to laser attacks according to claim 5, characterized in that: The construction process of the dynamic randomization computation path in step S5 includes the following steps: S61: During the chip design phase, the arithmetic logic units, registers, and memory modules involved in secure computing are physically divided into multiple logical computing clusters. S62: Before each secure computation is initiated, the security control unit randomly selects a configuration scheme from a preset set of legal path configurations based on the entropy source provided by a physically unclonable function or a ring oscillator. S63: The configuration scheme defines the physical order of the logical computing clusters through which the data flow passes in this calculation, the specific operation subtypes executed within each cluster, and the wiring switch status of the inter-cluster interconnection. S64: Dynamically establishes a physical computing path dedicated to this computing task based on the selected configuration scheme through a programmable interconnect network.

7. The method for designing a secure key storage and computation area for AI chips resistant to laser attacks according to claim 6, characterized in that: The proactive defense response mechanism in step S6 specifically includes the following steps: S71: Immediately terminate all currently running secure computing tasks and send an emergency wipe command to the temporary computing area and secure storage unit; S72: Disconnect the holding voltage to the non-volatile memory cells containing the critical key fragments and apply a high-voltage erase pulse to accelerate data annihilation; S73: Encrypt the attack characteristics, strength, and timestamp information identified by the multi-sensor fusion attack determination model and write them into an append-only audit log storage. S74: Based on the fusion attack probability value obtained in step S3 Based on the identified attack characteristics, dynamically adjust the parameters of the tunable filter in the multilayer heterogeneous shielding structure or the strength of the physical obfuscation mechanism used in step S2.

8. The method for designing a secure key storage and computation area for AI chips resistant to laser attacks according to claim 7, characterized in that: The dynamic adjustment of physical security parameters in step S74 specifically includes the following calculation steps: S741: Based on the identified attack characteristics, find the most sensitive blocking or obfuscation parameters for this type of attack from a preset mapping table. and its basic adjustment coefficient β; S742: Calculate the magnitude factor for this adjustment. Its value is related to the probability value of fusion attack. Positive correlation; S743: Set the current parameter Adjusting towards enhanced protection yields new parameter values. ; S744: Apply new parameter values The baseline vector on which the environmental safety assessment in step S342 depends is updated to the tunable filter or the confusion logic circuit in step S2, to form a feedback adjustment closed loop for the protection strength.

Images