Privacy computing protection method and system of AI large model service

Abstract

The application discloses an AI large model service privacy calculation protection method and system, relates to the technical field of artificial intelligence and data security, and comprises the following steps: in the privacy calculation protection running process, the calling time sequence and the privacy consumption decreasing record in the high concurrency stage are collected, alignment processing is carried out according to a unified time scale, a consumption change sequence is formed, and an exhaustion approximation mark is written at the tail of the consumption change sequence. Through the construction of the consumption change sequence and the exhaustion approximation mark, the combination of time density comparison and output precision corresponding analysis, the identification of the privacy consumption and the output rhythm dislocation section, and the dispersion of the concentrated decreasing process through rhythm rearrangement, precision decreasing transition and time scale reverse offset, the application eliminates the risk of continuous high-precision release in the critical stage from the time structure level, and improves the dynamic adjustment ability and overall security of the privacy calculation protection in the large model service scene.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and data security technology, specifically to a method and system for protecting privacy computation in AI large model services. Background Technology

[0002] Privacy-preserving computation protection for large AI model services refers to a comprehensive technical system that embeds privacy protection mechanisms throughout the entire computation process. This is crucial to prevent the leakage, misuse, or inference through reverse engineering of user input data, business data, and model internal parameters during processing, transmission, and storage, when large AI models provide training, fine-tuning, inference, and question-answering services via online interfaces, cloud inference, or platform capabilities. Built around the principle of "data usable but not visible," this system performs anonymization and permission verification before data enters the model. During the model inference phase, it introduces technologies such as encrypted computation, secure multi-party computation, and trusted execution environments, enabling collaborative computation between multiple parties without exposing plaintext data. At the result output stage, it implements privacy leakage risk assessment and content filtering, and audits and controls access to call behavior. This approach maintains model service capabilities while reducing risks such as data leakage, model inversion attacks, and member inference attacks, supporting data security and compliant operation in large model service scenarios.

[0003] The existing technology has the following shortcomings: In existing technologies, privacy-preserving computations for large-scale model services typically employ a cumulative deduction mechanism for privacy consumption values ​​to control the number of data calls, and limit high-frequency access behavior by setting protection thresholds. However, during periods of sudden traffic surges, requests flood in within a very short time, causing privacy consumption values ​​to be rapidly depleted or even instantly exhausted. Due to response delays or statistical lags in threshold determination in existing technologies, the system may continue to return high-precision results according to its normal output rhythm, easily leading to the continuous release of sensitive data without timely triggering of protection thresholds. Under conditions of high-frequency continuous calls, attackers can easily reconstruct core business data on a large scale in a short period of time through result splicing and reverse calculation, thus creating a serious risk of data leakage.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a privacy computing protection method and system for AI large model services, so as to solve the problems in the background art mentioned above.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for protecting privacy computing in AI large model services, comprising the following steps: During the operation of privacy-preserving computation, the call time series and privacy consumption reduction records during the high-concurrency phase are collected, aligned according to a unified time scale, and a consumption change sequence is formed. An exhaustion approach marker is written at the end of the consumption change sequence. Based on the consumption change sequence corresponding to the exhaustion approximation marker, time density comparison is performed to locate the time segment within a unit time scale where the privacy consumption deceleration rate exceeds a preset density threshold, forming a consumption transition trajectory, and marking the exhaustion critical position at the end of the consumption transition trajectory. Around the depletion critical position, the output accuracy record at the same time scale is continuously compared with the consumption change sequence to identify the time segment in the vicinity of the depletion critical position where the output accuracy remains at a high level, and an over-release mark is formed at the end of the corresponding time segment. Based on the over-release flag, trace back the corresponding consumption transition trajectory, locate the time misalignment segment where the decreasing rhythm of the consumption change sequence lags behind the rhythm of the output precision change, generate a rhythm misalignment segment record, and write the adjustment entry position at the end of the rhythm misalignment segment record. By adjusting the rhythm misalignment segment corresponding to the entry point, the rhythm of subsequent call time series is rearranged. A transition with decreasing output precision is introduced in the early part of the consumption transition trajectory, and the call time interval is extended in the region near the depletion critical position. The privacy consumption reduction concentration segment is dispersed by the reverse offset of the time scale, so as to complete the dynamic adjustment of privacy computing protection.

[0007] Preferably, the steps for collecting call time series and privacy consumption decreasing records during high-concurrency phases and aligning them according to a unified time scale are as follows: Collect call time series and privacy consumption reduction records during high-concurrency phases, establish a pairing relationship between the occurrence time and the occurrence time of reduction around the call identifier, and limit the continuous segments of the high-concurrency phase; Construct a unified time scale axis, map the call time series and privacy consumption reduction records to the corresponding scale points of the unified time scale axis, forming a consumption change sequence that includes the consumption value before reduction, the consumption value after reduction, the reduction magnitude, and the call count; Extract the remaining consumption based on the consumption change sequence and generate approximation events and exhaustion events based on the preset exhaustion baseline and preset approximation interval. Write an exhaustion approximation flag containing the reference time scale and the consumption value after the reference decrease at the end of the consumption change sequence. Establish a binding relationship between the exhaustion approximation identifier and the time scale index of the consumption change sequence, and attach a time granularity and mapping rules with a unified time scale to form an alignment result for subsequent steps.

[0008] Preferably, the consumption margin is expressed as the difference between the decreased consumption value and the preset depletion baseline, and the minimum consumption margin representative value of the continuous scale segment corresponding to the approximation event is used. The depletion approximation identifier includes the reference time scale, the reference consumption margin and the reference decrease magnitude, and establishes a corresponding relationship with the time scale index of the consumption change sequence.

[0009] Preferably, the steps for forming the exhaustion transition trajectory and marking the depletion critical position at the end of the exhaustion transition trajectory are as follows: A temporal density comparison benchmark is established around the consumption change sequence of the write exhaustion approximation flag. The decrease rate is extracted and a privacy consumption deceleration rate expression value is formed within a unit time scale. The privacy consumption deceleration rate expression value within a unit time scale is compared with the preset density threshold scale by scale. Based on the privacy consumption deceleration rate expression value within a unit time scale, extract time segments that continuously exceed a preset density threshold, and record the change trend of consumption value before and after the deceleration within the time segment; Consumption transition trajectories are formed by connecting time segments in a unified time scale sequence, generating trajectory entries that include the consumption value before and after the decrease. Read the reference time scale in the exhaustion approximation marker, compare the trajectory termination point of the exhaustion jump trajectory with the reference time scale at intervals, and mark the exhaustion critical position at the end of the exhaustion jump trajectory.

[0010] Preferably, the time density comparison reference plane uses a unified time scale as the horizontal main axis and the decreasing amplitude as the vertical expression. The privacy consumption deceleration rate expression value within a unit time scale is formed by the decreasing amplitude corresponding to the time scale length. The consumption transition trajectory uses the consumption value before the decrease at the trajectory starting scale point as the trajectory starting consumption reference and the consumption value after the decrease at the trajectory ending scale point as the trajectory ending consumption state.

[0011] Preferably, the over-release identification steps are as follows: Establish a time range for the neighboring region around the depletion critical position and map the output accuracy record to a unified time scale, forming a one-to-one correspondence between the time scale and the consumption change sequence; Based on continuous correspondence comparison between output accuracy records and consumption change sequences within the time range of neighboring regions, candidate time segments are extracted that correspond to the consumption margin range when the consumption value is at the depletion critical position after decreasing and the output accuracy level remains at a preset high precision state. Candidate time segments are screened by combining the depletion critical position, and the time segments that meet the conditions before the depletion critical position are determined and the changes in the consumption value after the decrease and the expression value of the privacy consumption deceleration rate within the unit time scale are recorded. Write an over-release flag containing the time scale difference on a unified time scale axis around the end point of the time segment that meets the conditions.

[0012] Preferably, the adjacent time range extends forward and backward around the depletion critical position according to a uniform time scale to form a fixed length interval. The candidate time segment is determined by the condition that the output result accuracy level continuously maintains a preset high accuracy state and the consumption value after reduction is located in the consumption margin interval corresponding to the depletion critical position. The over-release indicator includes the starting scale point, the ending scale point, and the time scale difference between the time segment and the depletion critical position.

[0013] Preferably, the steps for generating rhythm misalignment segment records and writing the adjustment entry position at the end of the rhythm misalignment segment records are as follows: Based on the over-release flag, extract the corresponding time segment and perform time scale backtracking in the consumption jump trajectory set to determine the consumption jump trajectory corresponding to the over-release flag and establish a time scale mapping relationship; By comparing the rhythm of the consumption jump trajectory and the time segment corresponding to the excess release indicator, candidate segments of rhythm misalignment are extracted where the consumption value change sequence after the decrease is not synchronized with the output result precision level change sequence. Based on the candidate segments of rhythm misalignment, the rhythm misalignment segment boundary is determined by splicing and generating a record of rhythm misalignment segments that includes the range of consumption value after decrease and the path of change of the precision level of the output result. By comparing the end point of the rhythm misalignment segment with the end point of the over-release indicator time segment, the adjustment entry position is written at the end of the rhythm misalignment segment record.

[0014] Preferably, the call time series is rearranged around the adjustment entry position, a decreasing output precision transition is introduced in the early part of the consumption transition trajectory, the call time interval is extended in the region near the depletion critical position, and the privacy consumption reduction steps are dispersed by reverse offset of the time scale as follows: Extract the continuous call time series after the adjustment entry position around the rhythm misalignment segment corresponding to the adjustment entry position, and arrange them side by side with the consumption change sequence on a unified time scale axis to form a rearranged preparation record containing call count, decrease magnitude and consumption value after decrease. The output precision records are rearranged around the first part of the consumption transition trajectory, and a decreasing output precision transition is introduced while the change in consumption value after the decrease is recorded synchronously. The call time series is reset at intervals around the region near the exhaustion critical position, the call records are redistributed to adjacent time scale points and the decrease magnitude is updated; By reversing the execution time scale of the call record around the adjustment entry position, the consumption value before and after the decrease is reconstructed, and a dynamic adjustment result is formed at the end of the record in the rhythm misalignment section.

[0015] The privacy-preserving computation system for AI large-scale model services includes a consumption sequence construction module, a transition trajectory recognition module, an over-release detection module, a rhythm misalignment localization module, and a rhythm rearrangement adjustment module. The consumption sequence construction module collects the call time sequence and privacy consumption decrease record during the high-concurrency phase of privacy computing protection operation, aligns them according to a unified time scale, forms a consumption change sequence, and writes an exhaustion approaching identifier at the end of the consumption change sequence. The transition trajectory recognition module compares the time density of the consumption change sequence corresponding to the depletion approximation marker, locates the time segment within a unit time scale where the privacy consumption deceleration rate exceeds a preset density threshold, forms a consumption transition trajectory, and marks the depletion critical position at the end of the consumption transition trajectory. The over-release detection module continuously compares the output accuracy records at the same time scale with the consumption change sequence around the depletion critical position, identifies the time segment in the vicinity of the depletion critical position where the output accuracy remains at a high level, and forms an over-release mark at the end of the corresponding time segment. The rhythm misalignment positioning module traces back the corresponding consumption transition trajectory based on the over-release flag, locates the time misalignment segment where the decreasing rhythm of the consumption change sequence lags behind the changing rhythm of the output precision, generates a rhythm misalignment segment record, and writes the adjustment entry position at the end of the rhythm misalignment segment record. The rhythm rearrangement adjustment module rearranges the rhythm of subsequent call time series around the rhythm misalignment segment corresponding to the adjustment entry position. It introduces a decreasing output precision transition in the early part of the consumption transition trajectory, extends the call time interval in the vicinity of the depletion critical position, and disperses the concentrated segment of privacy consumption decrease by offsetting the time scale in the opposite direction, thus completing the dynamic adjustment of privacy computing protection.

[0016] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention constructs a consumption change sequence and introduces a depletion approximation marker. Combined with a time density comparison and consumption transition trajectory recognition mechanism, it achieves continuous characterization and early perception of the privacy consumption rhythm, enabling a fine-grained expression of the concentrated decrease in privacy consumption during high-concurrency phases in the time dimension. Based on this, by continuously comparing the depletion critical position with the output accuracy record, it can identify time segments that maintain high-precision output within the critical proximity region. This eliminates the time misalignment window between privacy consumption decrease and output rhythm, avoiding the data exposure risk caused by continuously releasing high-precision results during the budget critical stage.

[0017] This invention generates rhythm misalignment segment records and sets adjustment entry positions to perform rhythm rearrangement on subsequent call time series. It introduces a decreasing output precision transition in the early stage of the consumption transition trajectory, extends the call time interval in the vicinity of the depletion critical position, and disperses the concentrated segment of privacy consumption decrease by reverse offset of the time scale. This allows the privacy consumption rhythm and output rhythm to be rematched under a unified time scale, weakening the splicing and calculation capability under continuous call conditions from the time structure level, and improving the proactive adjustment capability of privacy computing protection and the overall operational security in large model service scenarios. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0019] Figure 1 A flowchart illustrating the privacy-preserving computation method for the AI ​​large model of this invention.

[0020] Figure 2 A schematic diagram of the privacy computing protection system serving the AI ​​large model of this invention. Detailed Implementation

[0021] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0022] This invention provides, for example Figure 1 The privacy-preserving computation method for the AI ​​large model service shown includes the following steps: During the operation of privacy-preserving computation, the call time series and privacy consumption reduction records during the high-concurrency phase are collected, aligned according to a unified time scale, and a consumption change sequence is formed. An exhaustion approach marker is written at the end of the consumption change sequence. To enable a continuous characterization of the privacy consumption evolution process during high-concurrency phases and to provide a consistent semantic benchmark for subsequent rhythm analysis, the specific implementation steps are as follows: During the operation of privacy-preserving computation, time information directly bound to the call behavior is extracted and organized into a call time series, encompassing the entire process from service call entry to output. Simultaneously, privacy consumption decrease records directly associated with the same call behavior are extracted and organized into a decrease record sequence. The call time series records the occurrence time of each call in a monotonically increasing time representation, and the record content includes at least one of the following three types of time points: call identifier, occurrence time, request entry time, processing start time, and output return time. The privacy consumption decrease record records the change in privacy consumption value based on the decrease occurrence time and decrease magnitude, and the record content includes at least three of the following four elements: consumption identifier, decrease occurrence time, consumption value before decrease, consumption value after decrease, and decrease magnitude. To ensure that the call time series and privacy consumption decrease records can directly correspond under the same time reference, the occurrence time of the call time series is generated using a unified time source, and the decrease occurrence time of the privacy consumption decrease record is generated using the same time source. The system is designed to reorder records with inconsistent order during the collection phase, creating a one-to-one traceable pairing between the call time series and the privacy consumption reduction records. This pairing uses the call identifier as a link, mapping the occurrence time and the reduction occurrence time to the same call identifier, thus maintaining the stability of the correspondence between the call time series and the privacy consumption reduction records even during high-concurrency phases. Furthermore, the high-concurrency phase is defined as the time range within which the call identifier growth rate reaches a preset trigger condition per unit of time. The trigger condition is expressed using a fixed trigger rule: when the call identifier count meets a threshold constraint within multiple consecutive time scales, the system enters the high-concurrency phase. Once in the high-concurrency phase, the system continuously collects call time series and privacy consumption reduction records until exiting the high-concurrency phase. Exiting the high-concurrency phase uses a fixed trigger rule symmetrical to the entry condition: when the call identifier count falls back below the threshold constraint within multiple consecutive time scales, the system exits the high-concurrency phase. This results in continuous segments of call time series and privacy consumption reduction records specific to the high-concurrency phase.

[0023] Based on the generated high-concurrency phase call time series and privacy consumption decrease records, alignment processing is performed according to a unified time scale to form a consumption change sequence. The unified time scale adopts a preset time granularity definition, which is set to a fixed interval based on the time unit output by the time source. The fixed interval can be a millisecond-level fixed interval and remains unchanged within a single high-concurrency phase. The alignment processing establishes a time scale axis with the unified time scale as the main axis. The time scale axis increases at fixed intervals from the start time of the high-concurrency phase to the end time of the high-concurrency phase. Then, the occurrence time of each call record in the call time series is mapped to the nearest upper bound tick point of the time scale axis and written into the call count slot of the tick point. The decrease occurrence time of each decrease record in the privacy consumption decrease record is mapped to the nearest upper bound tick point of the same time scale axis and written into the decrease amplitude slot of the tick point. When there are multiple decrease records at the same tick point, the decrease amplitude slot is filled with the accumulated value of the decrease amplitude, while retaining the boundary value record between the consumption value before and after the decrease, so that the same tick point represents a single occurrence. The overall decrease within the time interval; when no decrease record appears at a certain scale point, the decrease amplitude slot is written to zero, and the consumption value before and after the decrease is taken from the consumption value after the decrease at the previous scale point, so that the consumption value is continuously expressed on the time scale without any gaps; when a call record appears at a certain scale point but no decrease record appears, the call count slot maintains the call count and synchronously writes a zero value decrease amplitude, so that the call behavior and consumption change remain aligned at the same scale point; when a decrease record appears at a certain scale point but no call record appears, the call count... The slots are written to zero while retaining the decrement range slots, ensuring that the concentrated deduction behavior remains visible on the time scale. After completing the above mapping, the consumption value before decrement, consumption value after decrement, decrement range, and call count for each scale point are concatenated into a time-seriesd entry according to the time scale. The time-seriesd entries are chained together in the order of the scale points to form a consumption change sequence. The consumption change sequence is indexed by the time scale, and the index entries simultaneously carry the call behavior intensity and privacy consumption decrement results, ensuring that subsequent steps can directly perform cross-record comparisons around the same time scale.

[0024] The consumption change trend is continuously annotated around the consumption change sequence to form the basis for writing the exhaustion approach indicator, which is then written at the end of the consumption change sequence. The basis for writing the exhaustion approach indicator is expressed using the consumption margin, which is the difference between the decreasing consumption value and the preset exhaustion baseline. The preset exhaustion baseline is the consumption state corresponding to the zero consumption margin. To avoid the impact of instantaneous jitter on the stability of the indicator, the consumption margin is expressed in continuous scale segments on the consumption change sequence, and the minimum consumption margin of the continuous scale segment is used as the representative value of the scale segment consumption margin. The length of the continuous scale segment is a fixed number of scales and remains unchanged within a single high-concurrency phase. When the representative value of the scale segment consumption margin enters the preset approximation interval, the preset approximation interval is expressed as the consumption margin falling into a fixed numerical range. The upper bound of the fixed numerical range is the preset approximation upper bound, and the lower bound is the preset exhaustion baseline. After setting the approximation interval, an approximation event is recorded in the consumption change sequence, and the termination scale of the approximation event is continuously updated. The termination scale of the approximation event is taken as the last scale point of the latest scale segment. When the representative value of the consumption balance of the scale segment reaches the preset exhaustion baseline, the approximation event turns into an exhaustion event and the exhaustion scale point is recorded. The exhaustion scale point is taken as the first scale point that reaches the preset exhaustion baseline. At the end of the high concurrency phase, the termination scale point corresponding to the approximation event is used as the reference point at the end of the consumption change sequence. The exhaustion scale point when the exhaustion event exists is taken as the priority of the tail reference point, and an identifier entry is added to the end of the consumption change sequence. The identifier entry is written with the exhaustion approximation identifier. The exhaustion approximation identifier contains four types of elements: reference time scale, reference consumption balance, reference consumption value after reduction, and reference reduction magnitude. In this way, the exhaustion approximation state is fixed at a clear position at the end of the consumption change sequence.

[0025] After the exhaustion approximation identifier is written, the association between the exhaustion approximation identifier and the consumption change sequence is solidified, and the alignment result is output for subsequent steps. The solidification method is expressed by index binding, establishing a one-to-one binding relationship between the reference time scale of the exhaustion approximation identifier and the time scale index of the consumption change sequence. This allows the exhaustion approximation identifier to directly point to the corresponding scale point entry within the consumption change sequence and to the set of continuous scale segment entries before the exhaustion approximation identifier is formed. Simultaneously, the collection range of the high-concurrency phase call time series and privacy consumption reduction records, the time granularity of the unified time scale, the mapping rules used for alignment processing, the set of entry fields in the consumption change sequence, and the exhaustion approximation identifier are all included. The set of elements is appended to the end of the consumption change sequence in the form of textual parameter records. The appended content is arranged adjacent to the exhaustion approximation identifier and maintains the same time scale reference. This allows the consumption change sequence to reproduce the alignment process under the same textual semantics and maintain the positioning consistency of the exhaustion approximation identifier when it is read by subsequent stages. This completes the entire process of collecting the call time series and privacy consumption reduction records during the high-concurrency stage of privacy computing protection, aligning them according to the unified time scale, forming the consumption change sequence, and writing the exhaustion approximation identifier at the end of the consumption change sequence. This ensures the continuity and traceability of the consumption change sequence and exhaustion approximation identifier during the high-concurrency stage.

[0026] Based on the consumption change sequence corresponding to the exhaustion approximation marker, time density comparison is performed to locate the time segment within a unit time scale where the privacy consumption deceleration rate exceeds a preset density threshold, forming a consumption transition trajectory, and marking the exhaustion critical position at the end of the consumption transition trajectory. To continuously characterize the evolution of privacy consumption in the consumption change sequence and extract concentrated decreasing segments, the specific implementation steps are as follows: A temporal density comparison benchmark is established around the consumption change sequence formed in the steps and with the exhaustion approximation marker written at the end. The temporal density comparison benchmark uses a unified time scale as the horizontal axis, the decrease magnitude corresponding to each time scale as the vertical expression, and the difference between the consumption value before and after the decrease in the consumption change sequence as a direct expression of the privacy consumption deceleration rate within a unit time scale. Based on this, the consumption change sequence is expanded into continuous scale segments according to the time scale order. Each continuous scale segment contains a fixed number of time scale points, and the length of the continuous scale segment remains constant within the same high-concurrency phase. For each time scale point, the decrease magnitude is extracted and divided by the corresponding time scale length to form the privacy within a unit time scale. The privacy consumption deceleration rate is expressed as a value, and the privacy consumption deceleration rate values ​​within a unit time scale of multiple adjacent time scales are arranged in chronological order to form a rate subsequence. After the rate subsequence is constructed, it is compared with a preset density threshold on a scale-by-scale basis. The preset density threshold is expressed as a fixed value and remains consistent within the same high-concurrency stage. The preset density threshold is used to define the boundary for determining when the privacy consumption deceleration rate reaches a concentrated decreasing state within a unit time scale. In this way, the consumption change sequence has a continuous and comparable deceleration rate expression basis in the time density dimension, providing a unified reference for subsequently locating the time segment where the privacy consumption deceleration rate within a unit time scale exceeds the preset density threshold.

[0027] After constructing the expression value of the privacy consumption deceleration rate within a unit time scale, continuous segment extraction is performed around the rate subsequence. Specifically, starting from the first time scale point of the consumption change sequence, the expression value of the privacy consumption deceleration rate within a unit time scale is read sequentially. When the expression value of the privacy consumption deceleration rate within a unit time scale at a certain time scale point exceeds a preset density threshold, that time scale point is marked as a density trigger scale point, and the rate subsequence is read continuously from this density trigger scale point as the starting point. In subsequent time scale points, as long as the expression value of the privacy consumption deceleration rate within a unit time scale continuously exceeds the preset density threshold, adjacent time scale points are included in the same time segment. When the expression value of the privacy consumption deceleration rate within a unit time scale returns to normal, the process is repeated. When the density falls below a preset density threshold, the extension of the current time segment is terminated, and the set of continuous time scale points from the density trigger point to the termination point is defined as the time segment in which the privacy consumption deceleration rate exceeds the preset density threshold within a unit time scale. Then, the scanning continues backward in the consumption change sequence, and the above extraction process is repeated until the entire consumption change sequence is traversed. In each extracted time segment, the changing trends of the consumption value before and after the reduction corresponding to each time scale point in the segment are recorded to maintain the correspondence between the time segment and the evolution of the consumption margin. This ensures that each time segment reflects both the dense state of the privacy consumption deceleration rate exceeding the preset density threshold within a unit time scale and retains the complete background information of the consumption change sequence.

[0028] After obtaining time segments where the privacy consumption deceleration rate exceeds a preset density threshold within multiple unit time scales, each time segment is concatenated in chronological order to form a consumption transition trajectory. Specifically, consecutive time scale points within the same time segment are arranged in a uniform time scale order, and the deceleration amplitude, consumption value before deceleration, consumption value after deceleration, and privacy consumption deceleration rate expression value within a unit time scale corresponding to each time scale point are combined into trajectory entries. These trajectory entries are arranged sequentially in chronological order to form trajectory segments. When multiple time segments are adjacent on the time scale axis and the interval does not exceed one time scale, the trajectory segments of adjacent time segments are spliced ​​together to form a continuous consumption transition. The consumption transition trajectory covers a continuous, concentrated decrease process occurring at a unified time scale. When there is an interval of more than one time scale between time segments, they are defined as independent consumption transition trajectories and numbered separately. Within each consumption transition trajectory, the time scale order remains unchanged, and the consumption value before the decrease at the trajectory's starting scale point is used as the trajectory's starting consumption benchmark, while the consumption value after the decrease at the trajectory's ending scale point is used as the trajectory's ending consumption state. The difference between the trajectory's starting consumption benchmark and the trajectory's ending consumption state expresses the overall privacy consumption decrease corresponding to that consumption transition trajectory, thus enabling the consumption transition trajectory to possess both continuity and integrity in the time and consumption dimensions.

[0029] After each consumption transition trajectory is formed, critical positions are marked around the exhaustion approximation marker already written at the end of the consumption change sequence. Specifically, the reference time scale and the reference decremented consumption value are first read from the exhaustion approximation marker, and the scale point entry corresponding to the reference time scale is located in the consumption change sequence. Then, the time interval between the trajectory termination scale point and the reference time scale is determined. When the trajectory termination scale point is before the reference time scale and the interval between them does not exceed a preset critical interval range, the trajectory termination scale point is defined as the exhaustion critical position, and an exhaustion critical position marker entry is added to the end of the corresponding consumption transition trajectory. The marker entry includes the time scale corresponding to the exhaustion critical position, the decremented consumption value, and the unit time. The data includes the privacy consumption deceleration rate within the scale and the time difference between the depletion approach marker and the reference time scale. When multiple consumption transition trajectories meet the above time interval conditions, the consumption transition trajectory whose termination scale point is closest to the reference time scale is selected as the priority annotation object, and the remaining trajectories retain their original numbers and are recorded as ordinary consumption transition trajectories. In this way, the depletion critical position is marked at the end of the consumption transition trajectory, so that the consumption transition trajectory and the depletion approach marker establish a clear association under a unified time scale. The entire process of comparing the time density based on the consumption change sequence corresponding to the depletion approach marker, locating the time segment where the privacy consumption deceleration rate exceeds the preset density threshold within a unit time scale, forming the consumption transition trajectory, and marking the depletion critical position at the end of the consumption transition trajectory is made public.

[0030] Around the depletion critical position, the output accuracy record at the same time scale is continuously compared with the consumption change sequence to identify the time segment in the vicinity of the depletion critical position where the output accuracy remains at a high level, and an over-release mark is formed at the end of the corresponding time segment. To achieve a fine characterization of the relationship between output behavior and privacy consumption evolution in the vicinity of the depletion critical position, the specific implementation steps are as follows: Based on the established exhaustion transition trajectory and the marked exhaustion critical position at the end of the corresponding trajectory, a neighboring region is established around the exhaustion critical position, and the time scale of the output precision records is aligned. Specifically, the time scale corresponding to the exhaustion critical position is read as the center scale point. Using a unified time scale as a reference, a fixed number of time scale points are extended forward and backward, thus forming a neighboring region time range centered on the exhaustion critical position. This neighboring region time range maintains a consistent length within the same high-concurrency phase. Subsequently, the output precision records within this neighboring region time range are extracted. The output precision records are indexed by the time scale, with each time scale point corresponding to one output precision record entry. The output precision record entry includes at least two of the following four elements: output result precision level, output data expression granularity, output result coverage, and output result detail retention. To ensure that the output precision record and the consumption change sequence can be directly and continuously compared, the output precision record is mapped to the same time scale axis as the consumption change sequence according to a unified time scale. When there are multiple output behaviors at a certain time scale point, the highest level of the output result precision level within the same time scale point is extracted according to a preset priority order and written into the output precision record entry of that time scale point. This makes the output precision record form a continuous expression on a unified time scale and maintain a one-to-one correspondence with the consumption change sequence in the time dimension.

[0031] After aligning the output precision record with the consumption change sequence on the time scale, continuous correspondence comparisons are performed at each time scale point within the adjacent time range. Specifically, the first time scale point within the adjacent time range is used as the starting point. The decreased consumption value in the consumption change sequence at that time scale point, as well as the privacy consumption deceleration rate expression value within a unit time scale, are read. Simultaneously, the output precision level and output data granularity in the output precision record entry at the same time scale point are read. Then, progressing step by step according to the time scale order, the consumption change sequence entry and output precision record entry are read simultaneously at each time scale point, and the decreased consumption value and output precision level are recorded side-by-side to form a continuous correspondence comparison record. During the continuous correspondence comparison... During the process, when the consumption value after reduction is within the consumption margin range corresponding to the depletion critical position and the output result accuracy level still maintains the preset high accuracy state, the current time scale point is marked as a high accuracy continuous scale point, and adjacent high accuracy continuous scale points are connected in chronological order to form a candidate time segment; when the output result accuracy level changes and no longer maintains the preset high accuracy state, the extension of the current candidate time segment is terminated and the start and end scale points of the candidate time segment are recorded; through the above continuous correspondence comparison, the correspondence between the consumption value after reduction and the output result accuracy level within the time range of the adjacent region is fully presented under a unified time scale, and a continuous data basis is provided for identifying time segments in the region adjacent to the depletion critical position where the output accuracy maintains a high accuracy state.

[0032] After extracting candidate time segments, these segments are screened to determine those that meet the criteria. Specifically, the distance between each candidate time segment and the time scale of the exhaustion critical position is calculated, expressed as a time scale difference. If the ending point of a candidate time segment is before the exhaustion critical position and the time scale difference between it and the exhaustion critical position does not exceed a preset proximity range, the candidate time segment is determined to meet the criteria. If a candidate time segment crosses the exhaustion critical position, it is truncated into two sub-segments using the exhaustion critical position as the boundary, and only the sub-segment before the exhaustion critical position is retained. The time intervals that meet the conditions are defined as follows: When determining the time intervals that meet the conditions, the changes in the decreasing consumption value and the changes in the privacy consumption deceleration rate expression value within the unit time scale are statistically analyzed. These changes are then recorded in correspondence with the number of scales on the scale where the output accuracy level remains at a preset high precision state. This allows the time intervals that meet the conditions to form a three-dimensional correlation expression in the time dimension, consumption dimension, and output accuracy dimension. Through the above screening, the time intervals that meet the conditions are accurately defined as the time intervals where the output accuracy remains at a high precision state in the vicinity of the depletion critical position, thus providing a clear target for the subsequent formation of the excess release indicator.

[0033] After determining the time segment that meets the conditions, an excess release flag is generated at the end of the corresponding time segment. In specific implementation, the end point of the time segment that meets the conditions is used as the flag writing position. A flag entry is added to the unified time scale axis of the consumption change sequence and output precision record. The flag entry includes the start point of the time segment, the end point of the time segment, the range of consumption values ​​after decreasing within the time segment, the range of privacy consumption deceleration rate expression values ​​per unit time scale within the time segment, and the number of scales at which the output result precision level continuously maintains a preset high precision state. At the same time, the time scale difference between the time scale at the exhaustion critical position and the end point of the time segment that meets the conditions is recorded in the flag entry. This method characterizes the temporal misalignment between maintaining high output accuracy and the depletion critical position. When multiple time segments meet the conditions, an excess release identifier is written at the end of each time segment and numbered according to the time scale sequence, so that each excess release identifier can establish a one-to-one correspondence with the corresponding time segment through the number. In this way, continuous correspondence comparison around the depletion critical position is completed under a unified time scale framework, and an excess release identifier is formed at the end of the corresponding time segment. This establishes a close correlation between the time segment maintaining high output accuracy and the consumption change sequence and the depletion critical position in the textual expression, providing a clear data foundation for subsequent rhythm misalignment analysis around the excess release identifier.

[0034] Based on the over-release flag, trace back the corresponding consumption transition trajectory, locate the time misalignment segment where the decreasing rhythm of the consumption change sequence lags behind the rhythm of the output precision change, generate a rhythm misalignment segment record, and write the adjustment entry position at the end of the rhythm misalignment segment record. To achieve a precise depiction of the temporal misalignment between the output rhythm and the privacy consumption rhythm, and to establish a clear entry point for subsequent rhythm adjustments, the specific implementation steps are as follows: Based on the established over-release identifier and corresponding time segment, the consumption transition trajectory is traced back in a directional manner according to the over-release identifier. Specifically, the start and end points of the time segment recorded in the over-release identifier are read, and a unified time scale range corresponding to that time segment is extracted. Then, in the established set of consumption transition trajectories, the start and end points of each consumption transition trajectory are retrieved one by one using the time scale as an index. When the end point of a consumption transition trajectory is located before the time segment corresponding to the over-release identifier and has a continuous time scale connection with the start point of the time segment, the consumption transition trajectory is determined to be the one corresponding to the over-release identifier. Consumption transition trajectory; if multiple consumption transition trajectories satisfy the above time connection relationship, they are sorted according to the time scale difference between the trajectory termination scale point and the starting scale point of the over-release flag time segment, and the consumption transition trajectory with the smallest time scale difference is selected as the target consumption transition trajectory; after determining the target consumption transition trajectory, the consumption value before the decrease, the consumption value after the decrease, and the privacy consumption deceleration rate expression value within a unit time scale corresponding to each time scale point within the target consumption transition trajectory are extracted, and a correspondence is established with the number of scales in the over-release flag where the output result accuracy level continuously maintains a preset high precision state, so that the target consumption transition trajectory and the over-release flag form a one-to-one mapping relationship under a unified time scale.

[0035] After completing the backtracking and mapping of the target consumption transition trajectory, a rhythmic comparative analysis is conducted around the time segments corresponding to the target consumption transition trajectory and the excess release indicator to locate the time misalignment segment where the decreasing rhythm of the consumption change sequence lags behind the changing rhythm of the output precision. Specifically, on a unified time scale axis, the decreasing consumption value change sequence of the target consumption transition trajectory and the output result precision level change sequence in the output precision record are arranged side-by-side, and the temporal order relationship of the two types of changes is observed step-by-step along the time scale. When the output result precision level changes from a preset high-precision state to a non-high-precision state at a certain time scale point, and the decreasing consumption value remains within the consumption margin range corresponding to the depletion critical position at that time scale point and several subsequent time scale points, the value will be determined from that time scale point. The time scale range from the point where the decreasing consumption value leaves the consumption margin interval corresponding to the depletion critical position is defined as the rhythm misalignment candidate segment. Conversely, when the decreasing consumption value leaves the consumption margin interval corresponding to the depletion critical position first, but the output result accuracy level still maintains the preset high accuracy state, the time scale range from the point where the decreasing consumption value leaves the consumption margin interval corresponding to the depletion critical position to the point where the output result accuracy level changes is defined as the rhythm misalignment candidate segment. When identifying the rhythm misalignment candidate segment, the decreasing amplitude corresponding to each time scale point in the rhythm misalignment candidate segment, the privacy consumption deceleration rate expression value per unit time scale, and the output result accuracy level are recorded simultaneously, so that the rhythm misalignment candidate segment maintains a complete expression in the time dimension and the change dimension.

[0036] After extracting candidate rhythm misalignment segments, these segments are merged and confirmed to form rhythm misalignment segments. Specifically, candidate rhythm misalignment segments with continuous time scales and consistent changes are spliced ​​together to form extended rhythm misalignment segments. When the time interval between adjacent candidate rhythm misalignment segments is one time scale increment, they are considered part of the same rhythm misalignment segment and merged. After forming rhythm misalignment segments, the boundaries of each segment are determined. The boundary starting point is the time scale point where the change in the accuracy level of the output result and the subsequent decrease in the consumption value are not synchronized. The boundary endpoint is the time scale point preceding the time scale point where the two types of changes restore the time consistency relationship. Subsequently, a rhythm misalignment segment record is generated for each rhythm misalignment segment. The rhythm misalignment segment record includes the starting scale point of the rhythm misalignment segment, the ending scale point of the rhythm misalignment segment, the range of the decreasing consumption value within the rhythm misalignment segment, the path of the change in the accuracy level of the output result within the rhythm misalignment segment, and the path of the change in the expression value of the privacy consumption deceleration rate within a unit time scale within the rhythm misalignment segment. The rhythm misalignment segment records are numbered according to the time scale order, so that multiple rhythm misalignment segment records form a continuous expression on the same time axis.

[0037] After generating the rhythm misalignment segment record, the adjustment entry position is written at the end of the rhythm misalignment segment record. Specifically, based on the termination point of each rhythm misalignment segment, a comparison is made between the termination point of the target consumption jump trajectory and the termination point of the time segment of the over-release indicator. When the termination point of the rhythm misalignment segment is before the termination point of the time segment of the over-release indicator, the termination point of the rhythm misalignment segment is defined as the adjustment entry position; when the termination point of the rhythm misalignment segment is after the termination point of the time segment of the over-release indicator, the termination point of the time segment of the over-release indicator is defined as the adjustment entry position. Subsequently, adjustments are added to the end of the rhythm misalignment segment record. The entry point entry includes the time scale corresponding to the entry point, the reduced consumption value corresponding to the entry point, and the output accuracy level status within one time scale before and after the entry point. This provides a clear starting point for subsequent rhythm rearrangement around the rhythm misalignment segment corresponding to the entry point. Through the above method, the entire process of tracing back the corresponding consumption transition trajectory based on the over-release flag, locating the time misalignment segment where the decreasing rhythm of the consumption change sequence lags behind the rhythm of the output accuracy change, generating rhythm misalignment segment records, and writing the entry point at the end of the rhythm misalignment segment records is made public, making the rhythm misalignment relationship traceable and adjustable within a unified time scale framework.

[0038] By adjusting the rhythm misalignment segment corresponding to the entry point, the rhythm of subsequent call time series is rearranged. A transition with decreasing output precision is introduced in the early part of the consumption transition trajectory. The call time interval is extended in the region near the depletion critical position. The privacy consumption reduction concentration segment is dispersed by the reverse offset of the time scale, thus completing the dynamic adjustment of privacy computing protection. To proactively intervene in the rhythmic relationship between privacy consumption and output behavior after recording rhythmic misalignment segments and determining the adjustment entry point, the specific implementation steps are as follows: Around the rhythm misalignment segment corresponding to the adjustment entry position, a comprehensive reconstruction of the subsequent call time series is prepared. Specifically, the time scale corresponding to the adjustment entry position is used as the starting point for rearrangement. Continuous call time series after the adjustment entry position are extracted on a unified time scale axis and arranged side-by-side with the existing consumption change sequence, so that each time scale point simultaneously displays the call count, decrease magnitude, and decreased consumption value. Then, the decreased consumption value corresponding to the adjustment entry position is used as the new adjustment baseline consumption value, and the time scale difference between the starting and ending points of the rhythm misalignment segment recorded in the rhythm misalignment segment record is used as the rhythm offset reference. This rhythm offset reference is written into the rearrangement preparation record, giving the subsequent call time series an adjustable offset range in the time dimension. Based on this, the unified time scale remains unchanged; only the call arrangement order on the time scale is adjusted, ensuring that the rearrangement is completed within the existing time frame without altering the overall time scale, thus guaranteeing the continued validity of the correspondence with the consumption change sequence.

[0039] After completing the reconstruction preparation, a decreasing output precision transition is introduced at the beginning of the consumption transition trajectory to reduce the continuous high-precision output state before the concentrated decrease. In specific implementation, the time scale range between the starting point of the target consumption transition trajectory and the adjustment entry position is traced back. Within this time scale range, the output precision records are rearranged so that the continuous time scale points that originally maintained the preset high precision state gradually reduce the output result precision level in chronological order. The decreasing transition of the output result precision level is expressed by a fixed step difference. Each time scale point corresponds to a precision level step. As the time scale moves towards the adjustment entry position, the output result precision level decreases step by step in chronological order, so that the output result presents a continuous transition state in the time dimension. During the decreasing output precision transition, the decrease magnitude and the consumption value after the decrease at the corresponding time scale point are recorded simultaneously, so that the change in the output result precision level and the privacy consumption decrease rhythm form a pre-match relationship on the time axis. By introducing a decreasing output precision transition at the beginning of the consumption transition trajectory, the output behavior before the concentrated decrease has already entered the transition stage, thereby reducing the basic conditions for the generation of rhythm misalignment segments.

[0040] After completing the transition of decreasing output precision, the call interval is extended around the region near the exhaustion critical position to disperse the call density within a unit time scale. Specifically, the time range of the region corresponding to the exhaustion critical position is identified on a unified time scale axis, and the call time series is reset at intervals within this region. The interval reset method is to redistribute multiple call records originally located at the same time scale point to adjacent subsequent time scale points in sequence, so that the call count within a unit time scale is reduced and distributed across multiple time scale points. During the redistribution process, the order of the call time series is kept unchanged, and only the corresponding time scale index is changed, so that the call behavior is extended backward in the time dimension. After the call interval is extended, the decrease amplitude of the corresponding time scale point is updated synchronously, so that the decrease amplitude originally concentrated at a single time scale point is distributed to multiple time scale points, so that the privacy consumption deceleration rate expression value within a unit time scale presents a balanced distribution in the neighboring region. By extending the call interval in the region near the exhaustion critical position, the consumption change sequence no longer shows a concentrated decrease peak in the critical stage, further weakening the conditions for the formation of rhythm misalignment segments.

[0041] After completing the transition of decreasing output precision and extending the call interval, the privacy consumption reduction concentration section is distributed through reverse offset of the time scale, thus achieving dynamic adjustment of privacy computing protection. Specifically, based on the adjustment entry position, reverse offset is performed on call records within a certain number of time scale points after the adjustment entry position. Some call records are allocated to idle time scale points before the adjustment entry position. The amount of reverse offset is referenced to the privacy consumption reduction rate expression value within a unit time scale recorded in the rhythm misalignment section record, ensuring that the reduction amplitude of each time scale point after the offset does not exceed a predetermined expression range. During the reverse offset process, the logical order of the call records is maintained, and only their time scale markers are adjusted. The privacy consumption reduction behavior is shifted forward on the time axis, echoing the transition of decreasing output precision. After the reverse offset is completed, the consumption value before and after the reduction for the corresponding time period of the consumption change sequence is regenerated and arranged in correspondence with the updated output precision record. This ensures that the privacy consumption reduction rate expression value within a unit time scale remains continuous and dispersed before and after the adjustment entry position. Through the above reverse offset of the time scale, the originally concentrated privacy consumption reduction segment is split and distributed across multiple time scale points in the time dimension. This breaks the continuous condition formed by the rhythm misalignment segment, realizes the dynamic adjustment of privacy computing protection, and re-establishes the synchronization relationship between the output rhythm and the consumption rhythm within a unified time scale framework.

[0042] This invention constructs a consumption change sequence and introduces a depletion approximation marker. Combined with temporal density comparison and consumption transition trajectory recognition mechanisms, it achieves continuous characterization and early perception of privacy consumption rhythm, enabling a fine-grained expression of the concentrated decrease in privacy consumption during high-concurrency phases in the temporal dimension. Furthermore, by continuously comparing the depletion critical position with output accuracy records, it can identify time segments that maintain high-precision output within the critical proximity region. This eliminates the temporal misalignment window between privacy consumption decrease and output rhythm, avoiding the data exposure risk caused by continuously releasing high-precision results during the budget critical phase.

[0043] This invention generates rhythm misalignment segment records and sets adjustment entry positions to perform rhythm rearrangement on subsequent call time series. It introduces a decreasing output precision transition in the early stage of the consumption transition trajectory, extends the call time interval in the vicinity of the depletion critical position, and disperses the concentrated segment of privacy consumption decrease by reverse offset of the time scale. This allows the privacy consumption rhythm and output rhythm to be rematched under a unified time scale, weakening the splicing and calculation capability under continuous call conditions from the time structure level, and improving the proactive adjustment capability of privacy computing protection and the overall operational security in large model service scenarios.

[0044] This invention provides, for example Figure 2 The privacy-preserving computation system for the AI ​​large model service shown includes a consumption sequence construction module, a transition trajectory recognition module, an over-release detection module, a rhythm misalignment localization module, and a rhythm rearrangement adjustment module. The consumption sequence construction module collects the call time sequence and privacy consumption decrease record during the high-concurrency phase of privacy computing protection operation, aligns them according to a unified time scale, forms a consumption change sequence, and writes an exhaustion approaching identifier at the end of the consumption change sequence. The transition trajectory recognition module compares the time density of the consumption change sequence corresponding to the depletion approximation marker, locates the time segment within a unit time scale where the privacy consumption deceleration rate exceeds a preset density threshold, forms a consumption transition trajectory, and marks the depletion critical position at the end of the consumption transition trajectory. The over-release detection module continuously compares the output accuracy records at the same time scale with the consumption change sequence around the depletion critical position, identifies the time segment in the vicinity of the depletion critical position where the output accuracy remains at a high level, and forms an over-release mark at the end of the corresponding time segment. The rhythm misalignment positioning module traces back the corresponding consumption transition trajectory based on the over-release flag, locates the time misalignment segment where the decreasing rhythm of the consumption change sequence lags behind the changing rhythm of the output precision, generates a rhythm misalignment segment record, and writes the adjustment entry position at the end of the rhythm misalignment segment record. The rhythm rearrangement adjustment module rearranges the rhythm of subsequent call time series around the rhythm misalignment segment corresponding to the adjustment entry position. It introduces a decreasing output precision transition in the early part of the consumption transition trajectory, extends the call time interval in the vicinity of the depletion critical position, and disperses the concentrated segment of privacy consumption decrease by offsetting the time scale in the opposite direction, thus completing the dynamic adjustment of privacy computing protection.

[0045] The privacy computing protection method for AI large model services provided in this embodiment of the invention is implemented through the aforementioned privacy computing protection system for AI large model services. For details of the specific methods and processes of the privacy computing protection system for AI large model services, please refer to the embodiments of the aforementioned privacy computing protection method for AI large model services, which will not be repeated here.

[0046] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for protecting privacy-preserving computations in large AI model services, characterized in that, Includes the following steps: During the operation of privacy-preserving computation, the call time series and privacy consumption reduction records during the high-concurrency phase are collected, aligned according to a unified time scale, and a consumption change sequence is formed. An exhaustion approach marker is written at the end of the consumption change sequence. Based on the consumption change sequence corresponding to the exhaustion approximation marker, time density comparison is performed to locate the time segment within a unit time scale where the privacy consumption deceleration rate exceeds a preset density threshold, forming a consumption transition trajectory, and marking the exhaustion critical position at the end of the consumption transition trajectory. Around the depletion critical position, the output accuracy record at the same time scale is continuously compared with the consumption change sequence to identify the time segment in the vicinity of the depletion critical position where the output accuracy remains at a high level, and an over-release mark is formed at the end of the corresponding time segment. Based on the over-release flag, trace back the corresponding consumption transition trajectory, locate the time misalignment segment where the decreasing rhythm of the consumption change sequence lags behind the rhythm of the output precision change, generate a rhythm misalignment segment record, and write the adjustment entry position at the end of the rhythm misalignment segment record. Around the rhythm misalignment segment corresponding to the adjustment entry position, the rhythm of subsequent call time series is rearranged. A transition with decreasing output precision is introduced in the early part of the consumption jump trajectory. The call time interval is extended in the region near the depletion critical position. The privacy consumption reduction concentration segment is dispersed by the reverse offset of the time scale.

2. The privacy computing protection method for AI large model services according to claim 1, characterized in that, The steps for collecting call time series and privacy consumption decreasing records during high-concurrency phases and aligning them according to a unified time scale are as follows: Collect call time series and privacy consumption reduction records during high concurrency phases, establish a pairing relationship between the occurrence time and the occurrence time of reduction around the call identifier, and limit the continuous segments of the high concurrency phase; Construct a unified time scale axis, and map the call time series and privacy consumption decrease records to the corresponding scale points of the unified time scale axis to form a consumption change sequence; Extract the remaining consumption based on the consumption change sequence and generate approximation events and exhaustion events based on the preset exhaustion baseline and preset approximation interval. Write an exhaustion approximation flag at the end of the consumption change sequence. Establish a binding relationship between the exhaustion approximation identifier and the time scale index of the consumption change sequence, and attach a time granularity and mapping rules with a unified time scale to form an alignment result for subsequent steps.

3. The privacy computing protection method for AI large model services according to claim 2, characterized in that, The consumption margin is expressed as the difference between the decreased consumption value and the preset depletion baseline. The minimum consumption margin representative value of the continuous scale segment corresponding to the approximation event is used. The depletion approximation identifier includes the reference time scale, the reference consumption margin, and the reference decrease magnitude, and establishes a correspondence with the time scale index of the consumption change sequence.

4. The privacy computing protection method for AI large model services according to claim 2, characterized in that, The steps for generating the exhaustion transition trajectory and marking the depletion critical position at the end of the exhaustion transition trajectory are as follows: A temporal density comparison benchmark is established around the consumption change sequence of the write exhaustion approximation flag. The decrease rate is extracted and a privacy consumption deceleration rate expression value is formed within a unit time scale. The privacy consumption deceleration rate expression value within a unit time scale is compared with the preset density threshold scale by scale. Based on the privacy consumption deceleration rate expression value within a unit time scale, extract time segments that continuously exceed a preset density threshold, and record the change trend of consumption value before and after the deceleration within the time segment; Consumption transition trajectories are formed by connecting time segments in a unified time scale sequence, generating trajectory entries that include the consumption value before and after the decrease. Read the reference time scale in the exhaustion approximation marker, compare the trajectory termination point of the exhaustion jump trajectory with the reference time scale at intervals, and mark the exhaustion critical position at the end of the exhaustion jump trajectory.

5. The privacy computing protection method for AI large model services according to claim 4, characterized in that, The time density comparison reference plane uses a unified time scale as the horizontal main axis and the decreasing amplitude as the vertical expression. The privacy consumption deceleration rate expression value within a unit time scale is formed by the decreasing amplitude corresponding to the time scale length. The consumption transition trajectory uses the consumption value before the decrease at the trajectory starting scale point as the trajectory starting consumption reference and the consumption value after the decrease at the trajectory ending scale point as the trajectory ending consumption state.

6. The privacy computing protection method for AI large model services according to claim 4, characterized in that, The steps for identifying over-released assets are as follows: Establish a time range for the neighboring region around the depletion critical position and map the output accuracy record to a unified time scale, forming a one-to-one correspondence between the time scale and the consumption change sequence; Based on continuous correspondence comparison between output accuracy records and consumption change sequences within the time range of neighboring regions, candidate time segments are extracted that correspond to the consumption margin range when the consumption value is at the depletion critical position after decreasing and the output accuracy level remains at a preset high precision state. Candidate time segments are screened by combining the depletion critical position, and the time segments that meet the conditions before the depletion critical position are determined and the changes in the consumption value after the decrease and the expression value of the privacy consumption deceleration rate within the unit time scale are recorded. Write an over-release flag on a unified time axis around the end point of the time segment that meets the conditions.

7. The privacy computing protection method for AI large model services according to claim 6, characterized in that, The adjacent time range extends forward and backward around the depletion critical position according to a uniform time scale to form a fixed length interval. The candidate time segment is determined by the condition that the output result accuracy level continuously maintains a preset high accuracy state and the consumption value after reduction is located in the consumption margin interval corresponding to the depletion critical position. The over-release indicator includes the starting point of the time segment, the ending point of the time segment, and the time scale difference between the time segment and the depletion critical position.

8. The privacy computing protection method for AI large model services according to claim 6, characterized in that, The steps for generating rhythm misalignment segment records and writing the adjustment entry position at the end of the rhythm misalignment segment records are as follows: Based on the over-release flag, extract the corresponding time segment and perform time scale backtracking in the consumption jump trajectory set to determine the consumption jump trajectory corresponding to the over-release flag and establish a time scale mapping relationship; By comparing the rhythm of the consumption jump trajectory and the time segment corresponding to the excess release indicator, candidate segments of rhythm misalignment are extracted where the consumption value change sequence after the decrease is not synchronized with the output result precision level change sequence. Based on the candidate segments of rhythm misalignment, the rhythm misalignment segment boundary is determined and a rhythm misalignment segment record is generated. By comparing the end point of the rhythm misalignment segment with the end point of the over-release indicator time segment, the adjustment entry position is written at the end of the rhythm misalignment segment record.

9. The privacy computing protection method for AI large model services according to claim 8, characterized in that, The call time series is rearranged around the adjustment of the entry position. A decreasing output precision transition is introduced in the early part of the consumption transition trajectory. The call time interval is extended in the region near the depletion critical position. The privacy consumption reduction steps are dispersed by reverse offset of the time scale as follows: Extract the continuous call time series after the adjustment entry position around the rhythm misalignment segment corresponding to the adjustment entry position, and arrange them side by side with the consumption change sequence on a unified time scale axis to form a rearranged preparation record containing call count, decrease magnitude and consumption value after decrease. The output precision records are rearranged around the first part of the consumption transition trajectory, and a decreasing output precision transition is introduced while the change in consumption value after the decrease is recorded synchronously. The call time series is reset at intervals around the region near the exhaustion critical position, the call records are redistributed to adjacent time scale points and the decrease magnitude is updated; By reversing the execution time scale of the call record around the adjustment entry position, the consumption value before and after the decrease is reconstructed, and a dynamic adjustment result is formed at the end of the record in the rhythm misalignment section.

10. A privacy computing protection system for AI large model services, used to implement the privacy computing protection method for AI large model services as described in any one of claims 1-9, characterized in that, It includes a consumption sequence construction module, a transition trajectory recognition module, an over-release detection module, a rhythm misalignment localization module, and a rhythm rearrangement adjustment module: The consumption sequence construction module collects the call time sequence and privacy consumption decrease record during the high-concurrency phase of privacy computing protection operation, aligns them according to a unified time scale, forms a consumption change sequence, and writes an exhaustion approaching identifier at the end of the consumption change sequence. The transition trajectory recognition module compares the time density of the consumption change sequence corresponding to the depletion approximation marker, locates the time segment within a unit time scale where the privacy consumption deceleration rate exceeds a preset density threshold, forms a consumption transition trajectory, and marks the depletion critical position at the end of the consumption transition trajectory. The over-release detection module continuously compares the output accuracy records at the same time scale with the consumption change sequence around the depletion critical position, identifies the time segment in the vicinity of the depletion critical position where the output accuracy remains at a high level, and forms an over-release mark at the end of the corresponding time segment. The rhythm misalignment positioning module traces back the corresponding consumption transition trajectory based on the over-release flag, locates the time misalignment segment where the decreasing rhythm of the consumption change sequence lags behind the changing rhythm of the output precision, generates a rhythm misalignment segment record, and writes the adjustment entry position at the end of the rhythm misalignment segment record. The rhythm rearrangement adjustment module rearranges the rhythm of subsequent call time series around the rhythm misalignment segment corresponding to the adjustment entry position. It introduces a decreasing output precision transition in the early stage of the consumption transition trajectory, extends the call time interval in the vicinity of the exhaustion critical position, and disperses the concentrated segment of privacy consumption decrease by offsetting the time scale in the opposite direction.

Images