Data forgetting method, electronic device, and computer-readable storage medium

Abstract

This application discloses a data forgetting method, electronic device, and computer-readable storage medium, relating to the field of data processing technology. The method includes: in response to a data deletion request, determining the speech record to be deleted; using a training slice associated with the speech record to be deleted as a target training slice, determining a target sub-model trained on the target training slice from the sub-models of the ensemble model; removing the speech record to be deleted from the target training slice, using the remaining data as a retraining sample set, and retraining the target sub-model, wherein the sub-models other than the target sub-model in the ensemble model remain unchanged; aggregating the retrained target sub-model with the unchanged sub-models to generate a new ensemble model after forgetting. This application effectively reduces the computational overhead of model retraining while supporting selective deletion of speech records.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a data forgetting method, electronic device, and computer-readable storage medium. Background Technology

[0002] Voice interaction systems generate and store a large amount of user voice data during the service process, including raw audio, transcribed text, and intent recognition results. With the continuous improvement of data privacy protection regulations (such as GDPR), users' control over their voice data is receiving increasing attention, especially the requirement that systems support users in deleting their personal historical voice data.

[0003] Currently, voice interaction systems typically respond to user data deletion requests in two ways: one is to directly delete the user's original audio files and corresponding text records at the storage layer, making the data inaccessible; the other is to remove the entire batch of training data and then retrain the model after removing it, in order to eliminate the impact of the deleted data on the model parameters.

[0004] However, the above methods have significant shortcomings. Storage layer deletion cannot eliminate the residual impact of deleted data on the parameters of the trained model; the deleted data may still be indirectly "remembered" through the model parameters. While full retraining can eliminate the model's influence, it requires retraining the entire model for each deletion request, resulting in extremely high computational costs. For scenarios where users selectively delete single or small amounts of voice recordings, the above methods struggle to balance operational granularity and computational overhead. Therefore, how to support selective deletion of voice recordings while reducing the computational cost of model retraining has become a pressing technical problem. Summary of the Invention

[0005] The main objective of this application is to provide a data forgetting method, electronic device, and computer-readable storage medium, which aims to solve the technical problem of reducing the computational overhead of model retraining while supporting selective deletion of voice recordings.

[0006] To achieve the above objectives, this application provides a data forgetting method, which includes the following steps: In response to a data deletion request, identify the voice recordings to be deleted; The training slice associated with the voice record to be deleted is used as the target training slice, and the target sub-model trained by the target training slice is determined from the sub-models of the ensemble model. Remove the speech records to be deleted from the target training slice, and use the remaining data after deletion as the retraining sample set to retrain the target sub-model. In this case, the sub-models other than the target sub-model in the ensemble model remain unchanged. The retrained target sub-model is aggregated with the remaining sub-models to generate a new ensemble model after forgetting.

[0007] In addition, to achieve the above objectives, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the data forgetting method described above.

[0008] In addition, to achieve the above objectives, this application also provides a readable storage medium, which is a computer-readable storage medium storing a computer program that is executed by a processor to implement the steps of the above method.

[0009] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described above.

[0010] One or more technical solutions proposed in this application have at least the following technical effects: This embodiment of the application, based on the correspondence between the ensemble model and training data shards, first determines the training shard to which the voice recording to be deleted belongs upon receiving a user's data deletion request. Then, it identifies the target sub-model associated with that shard from the ensemble model, thereby limiting the subsequent operation scope to that specific sub-model. Subsequently, the record to be deleted is removed from the target training shard, and the target sub-model is retrained using the remaining data, while the other sub-models in the ensemble model remain unchanged. Since the retraining process only involves a single sub-model and its corresponding partial training data, there is no need to retrain the entire ensemble model, thus significantly reducing computational resource consumption and time costs. At the same time, the parameters of the retrained target sub-model are updated based on a sample set that does not contain the deleted data, ensuring that the forgotten voice recordings no longer leave inferable traces at the model parameter level, thus ensuring the actual effect of data forgetting. Finally, the retrained target sub-model is aggregated with the remaining unchanged sub-models, and the resulting new ensemble model eliminates the influence of the specified data while maintaining the original overall performance, enabling it to quickly respond to user deletion requests. This effectively reduces the computational overhead of model retraining while supporting selective deletion of voice recordings. Attached Figure Description

[0011] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0012] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating the first embodiment of the data forgetting method of this application; Figure 2 This is a schematic diagram of the original data deletion process involved in one embodiment of the data forgetting method of this application; Figure 3 This is a schematic diagram illustrating the process for determining whether forgetting was successful in one embodiment of the data forgetting method of this application. Figure 4 This is a schematic diagram of the automatic expiration process involved in one embodiment of the data forgetting method of this application; Figure 5 This is a schematic diagram of the device structure of the hardware operating environment involved in the data forgetting method in the embodiments of this application.

[0014] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0015] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Voice interaction systems generate records of raw audio, transcribed text, intent results, confidence levels, device context, and training purpose. Existing privacy panels typically only display a portion of the data or provide a full wipe option, making it difficult to support selective management by user based on date, context, intent, or individual records.

[0017] For data already used in model training, simply deleting records in the storage layer cannot eliminate their impact on model parameters. Retraining the entire model after each deletion would be prohibitively costly and time-consuming.

[0018] To address the aforementioned issues, the main solution of this application is as follows: In response to a data deletion request, determine the audio recording to be deleted; use the training slice associated with the audio recording to be deleted as the target training slice, and determine the target sub-model trained through the target training slice from the sub-models of the ensemble model; remove the audio recording to be deleted from the target training slice, and use the remaining data after removal as the retraining sample set to retrain the target sub-model, wherein the sub-models other than the target sub-model in the ensemble model remain unchanged; aggregate the retrained target sub-model with the unchanged sub-models to generate a new ensemble model after forgetting.

[0019] This application, based on the correspondence between the ensemble model and training data shards, first determines the training shard to which the voice recording to be deleted belongs upon receiving a user's data deletion request. Then, it identifies the target sub-model associated with that shard from the ensemble model, thus limiting subsequent operations to that specific sub-model. Subsequently, the record to be deleted is removed from the target training shard, and the target sub-model is retrained using the remaining data, while the other sub-models in the ensemble model remain unchanged. Since the retraining process only involves a single sub-model and its corresponding portion of training data, there is no need to retrain the entire ensemble model, significantly reducing computational resource consumption and time costs. Simultaneously, the parameters of the retrained target sub-model are updated based on a sample set that does not contain the deleted data, ensuring that forgotten voice recordings leave no inferable traces at the model parameter level, guaranteeing the actual effect of data forgetting. Finally, the retrained target sub-model is aggregated with the remaining unchanged sub-models, generating a new ensemble model that eliminates the influence of the specified data while maintaining the original overall performance. This allows for rapid response to user deletion requests, effectively reducing the computational overhead of model retraining while supporting selective deletion of voice recordings.

[0020] It should be noted that the execution subject of each embodiment of the data forgetting method of this application can be an electronic device with data processing and program running functions, such as a mobile phone, server, personal computer, tablet computer, smart wearable device, Internet of Things device, etc. The embodiments of the data forgetting method of this application do not impose specific limitations on this.

[0021] Based on this, this application proposes a data forgetting method according to a first embodiment. In this embodiment, the data forgetting method includes the following steps S10 to S40: Step S10: In response to the data deletion request, determine the voice recordings to be deleted; Users can view generated voice interaction records through the history browsing module, which supports filtering and searching by date, scenario type, intent category, keywords, or data usage. Each record presents information including audio playback, transcribed text, recognition intent, usage time range, and associated model usage descriptions, facilitating quick location of the target record for management.

[0022] Based on browsing and searching, users can select one or more records through the interface and initiate a selective deletion request. The request includes the identification information of the voice record to be deleted, such as the record ID (Identifier), timestamp, user account, or dialogue round number.

[0023] Upon receiving a deletion request, user authentication and permission verification are performed to ensure that the requester has the necessary management rights to the data. If the verification passes, the system retrieves matching voice recordings from the storage structure based on the identifier information in the request and identifies these as the voice recordings to be deleted.

[0024] During the deletion process, all associated data, including the original audio file, ASR (Automatic Speech Recognition) transcribed text, NLU (Natural Language Understanding) intent recognition results, confidence scores, timestamps, scene labels, and usage labels, are removed. Simultaneously, feature caches, search indexes, and corresponding items in recoverable backups are cleared to ensure complete data erasure at all storage levels. For batch deletion scenarios involving multiple records, data can be processed individually or in a unified batch.

[0025] In addition, metadata about deletion requests (such as request time and request source) can be recorded for subsequent auditing and tracing.

[0026] Step S20: The training slice associated with the voice record to be deleted is used as the target training slice, and the target sub-model trained by the target training slice is determined from the sub-models of the ensemble model. When voice interaction records are generated, a unified index structure is established for each record, including the original audio, ASR-transcribed text, NLU intent recognition result, confidence score, timestamp, scene label, and purpose label. Simultaneously, it is explicitly marked whether the record is included in the training set, analysis set, or personalized dataset to facilitate subsequent data usage tracking and segment management.

[0027] Based on this, all speech training data used to train the ensemble model is pre-divided into multiple non-overlapping training slices. Each slice contains several speech records and their corresponding transcribed text, intent labels, and other annotation information. The partitioning strategy can be based on user ID hashing, time intervals, or random uniform sampling to ensure that the data distribution of each slice is relatively independent.

[0028] The ensemble model consists of multiple sub-models, each trained independently using a training slice, thus establishing a mapping relationship between the sub-models and the training slices. After identifying the speech record to be deleted, the index information of each training slice is traversed to find the slice containing the record and mark it as the target training slice.

[0029] Based on the pre-stored sub-model-segment mapping table, the sub-model trained by the target segment is found and used as the target sub-model. The remaining sub-models are not directly related to the deletion request.

[0030] Step S30: Remove the speech records to be deleted from the target training segment, and use the remaining data after removal as the retraining sample set to retrain the target sub-model. In this case, the sub-models other than the target sub-model in the ensemble model remain unchanged. After determining the target training slice and target sub-model, remove all associated data of the speech records to be deleted from the target training slice. This may include the original audio file, transcribed text, acoustic features, intent annotations, etc. The deletion operation can be physical deletion or logical markup deletion, but it must be ensured that this data will not be read or used again during subsequent retraining.

[0031] The remaining data after removal is used as the retraining sample set to retrain the target sub-model. During retraining, the same network architecture, hyperparameters, and optimization algorithms as the original training can be used. The number of training epochs can be adjusted appropriately according to the amount of remaining data. Transfer learning or incremental learning strategies can also be used to accelerate convergence.

[0032] Throughout the retraining process, other sub-models in the ensemble model do not participate in the update, maintaining their original parameters and structure to preserve their learned knowledge.

[0033] Step S40: Aggregate the retrained target sub-model with each of the unchanged sub-models to generate a new ensemble model after forgetting.

[0034] After the target sub-model is retrained, it is combined with the remaining sub-models that remain unchanged to generate a new ensemble model after forgetting. Each sub-model maintains an independent parameter structure after combination. During the inference phase, the input data is fed into each sub-model separately, and each sub-model outputs its prediction result independently. Then, the output results are fused according to the set ensemble method to obtain the final prediction result.

[0035] The combined model maintains overall functionality identical to the original model, but eliminates the impact of deleted voice recordings on model parameters. The new integrated model is then deployed, replacing the old model, and its version history is updated, completing the entire data forgetting process.

[0036] This process supports repeated execution of single or batch deletion requests, and each operation only affects a single sub-model, ensuring the efficiency and scalability of the operation.

[0037] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. On this basis, refer to Figure 2 As shown, after the step of determining the voice recording to be deleted, the method further includes: Step A10: Obtain the record index of the voice recording to be deleted; After identifying the voice recording to be deleted, obtain the record index corresponding to that recording.

[0038] The record index is established when the voice interaction record is generated, assigning a unique index identifier to each record. This index identifier is mapped to all associated data of the record (including the original audio file, transcribed text, intent tags, timestamps, etc.). The record index can be implemented using hash values, UUIDs, or auto-incrementing IDs, and stored in an index table for fast retrieval.

[0039] Step A20: Query the storage location of the original data according to the record index. The original data includes the audio file and / or text file corresponding to the voice record to be deleted in the storage layer. Based on the obtained record index, query the storage location of the original data corresponding to the voice record to be deleted in the storage layer.

[0040] This storage location points to the path information of the original audio file and / or text file in the actual physical storage medium (such as a local disk, distributed file system, or object storage).

[0041] By mapping the index to the storage location, all the original data files associated with the record can be accurately located, ensuring that subsequent deletion operations can accurately hit the target.

[0042] Step A30: Delete the original data at the storage location.

[0043] At the queried storage location, perform the deletion operation on the original data. The deleted objects include the audio file and / or text file corresponding to the voice recording to be deleted.

[0044] Deletion methods can be physical deletion (i.e., completely erasing file data and freeing up storage space) or logical deletion (e.g., marking the file as inaccessible for later unified cleanup), but it must be ensured that the original data cannot be read or recovered after deletion.

[0045] This step is independent of and executed in parallel with the operation of removing the voice records to be deleted from the training slices in step S30 above. It completes the data forgetting process from both the storage layer and the model layer, ensuring that the original data is completely cleared at the storage layer.

[0046] In one possible implementation, prior to the step of using the training segment associated with the voice recording to be deleted as the target training segment, the method further includes: Step B10: Obtain the usage tag from the record index of the voice recording to be deleted; This usage tag is established and stored in the record index when the voice interaction record is generated. It is used to mark the data type of the record, which may include categories such as training set, analysis set, or personalized dataset.

[0047] By reading the usage tag, it is possible to quickly determine whether the voice recording to be deleted has been used for model training or other data analysis tasks, providing a basis for judgment on whether to perform model layer forgetting processing in the future.

[0048] Step B20, if the usage label indicates that the speech record to be deleted has been involved in model training, performs the step of using the training slice associated with the speech record to be deleted as the target training slice.

[0049] After obtaining the usage label, the label is identified and judged.

[0050] If the usage label indicates that the voice recording to be deleted has been used in model training (i.e., belongs to the training set), then proceed to step S20 to initiate the data forgetting process of the model layer.

[0051] If the usage label indicates that the voice recording to be deleted did not participate in model training (e.g., it only belongs to the analysis set or personalized dataset), then there is no need to perform the localization of the training segments and subsequent retraining operations. Only the deletion of the original data in the storage layer in the aforementioned steps A10 to A30 can be performed, thereby avoiding unnecessary computational overhead.

[0052] In one possible implementation, the step of aggregating the retrained target sub-model with the unchanged sub-models to generate a new ensemble model after forgetting includes: Step C10: Obtain the retrained target sub-model and each sub-model in the ensemble model other than the target sub-model; After retraining, the parameters of the target sub-model have been updated based on the remaining sample set after removing the speech records to be deleted. The updated model parameters are saved to obtain the retrained target sub-model for subsequent combination.

[0053] Step C20: Combine the retrained target sub-model with all other sub-models to form a new ensemble model after forgetting.

[0054] The combined new integrated model logically retains the independent identity of each sub-model. The sub-models do not undergo parameter mixing or joint optimization, but coexist in the same integrated framework in a parallel manner.

[0055] During the inference phase, the input data is fed into each sub-model of the new ensemble model. Each sub-model performs inference independently and outputs its own prediction results. Then, the output results of all sub-models are fused according to a predetermined ensemble method (such as voting, weighted voting, etc.) to obtain the final prediction output.

[0056] Once combined, the new integrated model can replace the original integrated model for subsequent voice interaction service inference. Since each sub-model remains independent, when a data deletion request for other voice records is received again, the corresponding individual sub-model can still be independently retrained and replaced according to the mapping relationship between the segment and the sub-model, without affecting other sub-models in the new integrated model.

[0057] In one possible implementation, refer to Figure 3 As shown, after the step of generating the new ensemble model after forgetting, the method further includes: Step D10: Obtain the feature vector of the voice recording to be deleted; This feature vector can be extracted and retained before data deletion, or it can be re-extracted during the verification phase based on the original audio of the voice recording to be deleted.

[0058] The feature vector extraction method must be consistent with the feature extraction method used in the training phase to ensure the effectiveness of subsequent confidence assessment.

[0059] Step D20: Input the feature vector into the new ensemble model to obtain the prediction confidence of the new ensemble model for the voice record to be deleted; Each sub-model in the new ensemble model infers the feature vector and outputs its own prediction result. The output results are then fused to obtain the final prediction confidence.

[0060] The prediction confidence level reflects the degree of certainty of the new ensemble model regarding the category of the input speech recording. If the impact of the speech recording to be deleted has been effectively eliminated, the performance of the new ensemble model on that data should be consistent with the performance of models that have not seen that data, i.e., it cannot give a high-confidence prediction.

[0061] Step D30: If the deviation between the prediction confidence level and the random guess level is less than or equal to a preset threshold, the forgetting is determined to be complete.

[0062] The prediction confidence level is compared with the random guess level, and the deviation between the two is calculated. The random guess level can be determined based on the total number of intent categories in the voice interaction system. For example, if the system has N intent categories, the baseline confidence value for random guessing is 1 / N. If this deviation is less than or equal to a preset threshold, it is determined that forgetting has been completed, meaning that the voice recording to be deleted no longer has a identifiable impact on the parameters of the new ensemble model.

[0063] If the deviation is greater than the preset threshold, it means that the new ensemble model still remembers the information of the data to a certain extent. At this time, the relevant hyperparameters in the retraining process can be adjusted (such as increasing the number of training rounds, adjusting the learning rate, or changing the batch size), and the retraining operation of the target sub-model in step S30 can be re-executed. Then, the forgetting verification in steps D10 to D30 can be executed again until the verification is passed.

[0064] To mitigate the impact of random fluctuations that may arise from a single validation, the prediction confidence level can be obtained multiple times, and the average value can be compared with the random guess level. If the deviation of the average value still exceeds a preset threshold, the aforementioned retraining adjustment operation is triggered. In the extreme case where the forgetting validation still fails after multiple adjustments to the retraining parameters, all training data in the target training segment can be used as an exclusion set. The target sub-model can then be completely retrained using all data in the original training set except for all data in the target segment, in order to achieve a more thorough forgetting effect.

[0065] The preset threshold can be set according to the actual needs of specific application scenarios, such as 0.05 or 0.1, or it can be adaptively adjusted according to the statistical characteristics of the confidence distribution. Furthermore, member inference detection can be used as an auxiliary verification method for forgetting completion. The feature vector of the voice record to be deleted is input into the new ensemble model, and the member inference attack method is used to detect whether the record was used for training. For example, the verification condition can be expressed as: |Amem(Ddel) 0.5|≤τ, where A_mem represents the member inference detection result (usually the probability value of the binary classification output), 0.5 represents the random guessing level, and τ represents the allowable deviation threshold. If this inequality holds, that is, the deviation between the detection result and the random guessing level does not exceed the allowable deviation threshold, it indicates that the new ensemble model can no longer effectively determine whether the record has participated in training, thus verifying the thoroughness of data forgetting from another perspective.

[0066] This verification condition can be used in conjunction with the aforementioned confidence bias condition to enhance the reliability of the forgetting determination. If multiple automatic remediation attempts fail the verification, the failure information of the forgetting operation can be recorded (including the identifier of the record to be deleted, the reason for failure, and the records of attempted retraining parameter adjustments, etc.), and a manual intervention review process can be triggered, with maintenance or compliance personnel handling the follow-up according to the specific circumstances.

[0067] In one possible implementation, refer to Figure 4 As shown, the data forgetting method further includes: Step E10: Obtain the retention period configured by the user; The retention period is set by the user based on their personal privacy management needs or compliance requirements, indicating the maximum time the user allows voice interaction records to remain in the storage system. Users can input a specific retention period value through the user interface or privacy settings module, such as setting it to 30 days, 90 days, or a custom duration.

[0068] The retention period can be configured uniformly for all voice recordings, or different retention periods can be set according to scene tags, usage tags, or data types (such as training sets, analysis sets, and personalized datasets) to meet users' differentiated retention needs for different types of data.

[0069] Step E20: Scan the voice records according to the retention period, and determine the voice records whose generation time and retention period sum to the current time as expired records; The voice recordings in the storage system are scanned periodically or continuously according to the retention period configured by the user.

[0070] The specific scanning method can be as follows: obtain the generation timestamp of each voice record, add the generation time to the retention period configured by the user, and obtain the predetermined expiration time of the record. When the expiration time is reached or earlier than the current system time, the voice record is determined to be an expired record.

[0071] The scanning process can be triggered uniformly according to a predetermined time cycle (e.g., executed at a fixed time every morning), or it can be triggered immediately after the retention period configuration is changed by an event. For users who have configured a retention period, all voice records under that user account will be scanned; for users who have not configured a retention period, the system's default retention period can be used or automatic expiration and deletion can be skipped.

[0072] After the scan is completed, all voice recordings that are determined to be expired are compiled into an expired record set, and the identification information and reason for expiration of each expired record are recorded for subsequent deletion operations.

[0073] Step E30: The expired record is used as the voice record to be deleted, and the step of using the training segment associated with the voice record to be deleted as the target training segment is executed.

[0074] For each record or batch of records in the expired record set, the training segment location operation in step S20 (i.e., finding the training segment to which the expired record belongs and determining the corresponding target sub-model) is executed in sequence. Then, the elimination and retraining operation in step S30 and the sub-model combination operation in step S40 are executed, and finally a new ensemble model that has forgotten the expired record is generated.

[0075] If there are multiple expired records in the same batch distributed across different training segments, the above operations can be performed sequentially on the target sub-models corresponding to each segment. Alternatively, multiple expired records in the same segment can be removed at once and then retrained to improve processing efficiency.

[0076] For expired records that only exist in the analysis set or personalized dataset but have never been used in model training, the model layer forgetting operation in steps S20 to S40 is unnecessary; only the original data deletion in the storage layer needs to be performed. Furthermore, before automatic deletion, a deletion reminder can be sent to the user via the notification module, informing them of the list of records expiring soon and the deletion time, allowing the user to choose to retain or export the relevant records in advance during the buffer period.

[0077] For example, to help understand the technical concept or principle of the data forgetting method combined with the first embodiment, a specific embodiment is now listed. In this specific embodiment, the data forgetting process includes: Step 1: Establish a voice recording index. When generating voice interaction records, the system establishes a unified index for the original audio, ASR-transcribed text, NLU intent, confidence level, timestamp, scene label, and purpose label, and records whether it has entered the training set, analysis set, or personalized dataset.

[0078] Step two provides historical browsing and retrieval. Users can filter records in the historical browsing module by date, scenario, intent, keywords, or data usage, and view the audio playback, transcribed text, recognition intent, usage time range, and associated model usage for each record.

[0079] Step 3: Receive selective deletion requests. After the user selects one or more records, the deletion execution module verifies the user's identity and authorization scope, and then deletes the corresponding items from the original audio, transcribed text, feature cache, search index, and recoverable backup.

[0080] Step four: Determine the scope of model impact. The usage mapping module queries the relevant training task, model version, and SISA training slice based on the record index. If the record has not been included in the training set, only data deletion is performed; if the record has participated in training, a model forgetting task is generated.

[0081] Step 5: Perform SISA shard isolation forgetting. The model forgetting module only retrains the shards containing the record to be deleted and re-aggregates them with the sub-models of unaffected shards to form a new model version. For cases where multiple models use the same record, the system generates forgetting tasks separately for each model version.

[0082] Step Six: Perform automatic expiration processing. The automatic expiration module scans records according to the retention period set by the user, such as 30 days, 6 months, 1 year, or permanent. Records that have reached their expiration time are entered into the deletion queue, and data deletion and necessary model obsolescence are completed according to Steps Three through Five.

[0083] Step 7: Verification and Auditing. The audit and verification module records the deletion request, processing time, affected shards, retraining results, and member inference detection results. When the verification conditions are met, the task is marked as completed; otherwise, it enters the review or retraining process.

[0084] Based on this, let's take user A deleting 10 voice recordings from May 14, 2025 as an example. The user requests the deletion of the original audio, the transcription of the text, and the elimination of the impact of these recordings on the trained ASR, TTS, or NLU models.

[0085] In this specific application scenario, the data forgetting process is as follows: Step 1: User A filters records from May 14, 2025 by date in the browsing history module, selects 10 records, and submits a deletion request.

[0086] Step two: The deletion execution module deletes the corresponding recordings, transcribed text, search indexes, and feature caches within approximately 50ms, and generates an irreversible deletion marker.

[0087] Step 3: The purpose mapping module found that the 10 records involve 3 out of 5 training slices and are associated with an ASR training task and an NLU training task.

[0088] Step 4: The model forgetting module only retrains the 3 affected segments, with approximately 5,000 samples per segment, for a total of approximately 15,000 samples; the unaffected segments remain unchanged.

[0089] Step 5: After retraining, the system re-synthesizes the model version and performs member inference detection. The memoization detection value of the records to be deleted decreased from approximately 0.92 to approximately 0.51, approaching the level of random guessing.

[0090] It should be noted that the above examples are only used to help understand this embodiment and do not constitute a limitation on the data forgetting process of this embodiment. Any simple modifications based on this technical concept are within the protection scope of this application.

[0091] Based on the first and / or second embodiments of this application, in the third embodiment of this application, the content that is the same as or similar to the first and second embodiments described above can be referred to the above description and will not be repeated hereafter. Following the step of generating the new ensemble model after forgetting, the method further includes: Step F10: Obtain the first prediction accuracy of the new ensemble model on the benchmark set and the second prediction accuracy of the ensemble model on the benchmark set, wherein the benchmark set contains speech samples of multiple intent categories; This benchmark set is a pre-built, standardized test dataset containing speech samples of multiple intent categories, and is guaranteed to have no overlap with the data in all training slices. It is used to objectively evaluate the generalization performance of the model. The newly ensemble model after forgetting is tested for inference on this benchmark set, and the proportion of samples that make correct predictions is recorded as the first prediction accuracy.

[0092] Obtain the second prediction accuracy of the original ensemble model on the same benchmark set.

[0093] The second prediction accuracy can be calculated and stored immediately after the original ensemble model is trained, or it can be obtained by reloading the original ensemble model for testing during the subsequent validation phase. By comparing the first prediction accuracy with the second prediction accuracy, the impact of the forgetting operation on the overall model performance can be evaluated.

[0094] Step F20: If the difference between the first prediction accuracy and the second prediction accuracy is greater than a preset threshold, it is determined that the target sub-model corresponding to the target training slice has undergone feature shift during retraining. If the difference exceeds a preset threshold, it indicates that the target sub-model experienced feature shift during retraining. Feature shift refers to the difference in data distribution between the remaining samples and the original complete samples after the removal of speech records to be deleted from the target training segment. This difference causes the retrained target sub-model to have a decreased predictive ability on samples belonging to the same intent category as the deleted samples, thus affecting the accuracy of the entire ensemble model on that category. The preset threshold can be set according to the tolerance for performance loss in the actual application, for example, it can be set to a value such as 0.01 or 0.02.

[0095] Step F30 triggers the compensation retraining of the target sub-model, wherein the compensation retraining uses samples in the benchmark set that belong to the same intent category as the speech record to be deleted as supplementary samples to perform incremental training on the target sub-model.

[0096] Samples belonging to the same intent category as the speech recordings to be deleted are selected from the benchmark set and added to the retraining sample set as supplementary samples. These supplementary samples are consistent with the speech recordings to be deleted in intent category, which can effectively compensate for the changes in the data distribution of this type caused by the deletion of records, thereby helping the target sub-model to restore its ability to distinguish this intent category while maintaining the forgetting effect.

[0097] Incremental training is performed on the target sub-model using the expanded sample set. This involves updating the parameters of the sub-model based on the already retrained parameters, rather than retraining from scratch, to reduce additional computational overhead. After the compensated retraining is completed, the accuracy verification steps F10 to F20 are performed again. If the difference is still greater than the preset threshold, the number of supplementary samples can be increased from the benchmark set or the hyperparameters of the incremental training can be adjusted before performing compensated retraining again, until the accuracy difference meets the requirements.

[0098] In one possible implementation, after the step of using the training segment associated with the voice recording to be deleted as the target training segment, the method further includes: Step G10: Obtain the total number of times the data deletion request was initiated within the preset time window; The preset time window can be pre-set according to the actual operation of the system. For example, it can be set to a time span of 24 hours, 7 days or 30 days, which is used to count the frequency of users initiating deletion requests within this time period.

[0099] The total number of deletion requests can be counted based on the deletion request log. Each valid deletion request is incremented by one, and the counter is automatically reset to zero and counted again after the time window expires.

[0100] Step G20: If the total number of times is greater than the preset frequency threshold, count the target training slices corresponding to each data deletion request to obtain the number of times each training slice is hit. A preset frequency threshold is used to determine whether the current deletion request is in a high-frequency state. For example, it can be set to exceed 10 times in 24 hours. If the total number of times does not exceed the threshold, it indicates that the deletion request is within the normal frequency range and no additional processing is required; if it exceeds the threshold, the target training slice located by each deletion request is recorded and accumulated.

[0101] This statistic can identify which training segments are frequently hit in a short period of time, thus reflecting the phenomenon of concentrated deletion of data records in certain specific segments.

[0102] Step G30: If the percentage of times any of the training segments is hit exceeds a preset threshold, a segment reconstruction instruction is generated, and retraining is performed on the sub-model corresponding to the training segment after data redistribution. The segment reconstruction instruction is used to redistribute the data in the first training segment to other training segments, where the first training segment is the training segment whose percentage of times is hit exceeds the preset threshold.

[0103] The percentage of times a deletion request is made refers to the proportion of times a specific training shard is hit out of the total number of deletion requests. This proportion reflects the concentration of deletion requests at the shard level. A preset percentage threshold can be set according to actual needs, such as 30% or 50%. When the percentage of times a shard is hit exceeds this threshold, it indicates that a large amount of data in that shard has been deleted in a concentrated manner, potentially resulting in insufficient remaining data in that shard and affecting the effectiveness of subsequent retraining.

[0104] The shard reconstruction command is used to trigger a data allocation adjustment operation, which means redistributing the remaining data in the first training shard (i.e., the training shard whose hit rate exceeds a preset threshold) to other training shards. Reconstruction must follow data allocation rules and can be performed using methods such as remapping based on user ID hashes, random uniform redistribution, or migration based on data distribution similarity.

[0105] After the data is redistributed, the sub-models corresponding to each redistributed training partition are retrained. Specifically, all sub-models corresponding to training partitions where the data has changed must be retrained based on the new partition data to ensure that each sub-model matches its corresponding new partition data. For sub-models corresponding to training partitions where the data has not changed, the original parameters are kept unchanged, and no retraining is required to control the overall computational cost.

[0106] Furthermore, embodiments of this application also propose an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the data forgetting method described above.

[0107] The electronic device provided in this application, employing the data forgetting method described in the above embodiments, can solve the technical problem of reducing the computational overhead of model retraining while supporting selective deletion of voice recordings. Compared with the prior art, the beneficial effects of the electronic device provided in this application are the same as those of the data forgetting method described in the above embodiments, and other technical features of this electronic device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0108] refer to Figure 5 It illustrates a structural schematic diagram of an electronic device suitable for implementing the embodiments of this application. The electronic device in the embodiments of this application may also include, but is not limited to, mobile phones, personal computers, servers, smart wearable devices, etc. Figure 5 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0109] like Figure 5 As shown, the electronic device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the electronic device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the system may be connected to the I / O interface 1006 to: input devices 1007, such as touchscreens, touchpads, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008, such as liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003, such as magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. While electronic devices with various systems are shown in the figures, it should be understood that implementation or possession of all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0110] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0111] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0112] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0113] In addition, to achieve the above objectives, embodiments of this application also provide a readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the data forgetting method described in the above embodiments.

[0114] The computer-readable storage medium provided in this application embodiment may be, for example, a USB flash drive, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0115] The aforementioned computer-readable storage medium may be included in an electronic device or may exist independently without being assembled into an electronic device.

[0116] The aforementioned computer-readable storage medium carries one or more programs that, when executed by an electronic device, cause the electronic device to perform the process steps of any embodiment of the data forgetting method.

[0117] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0118] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0119] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules are not, in some cases, limiting the scope of the module itself.

[0120] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described data forgetting method. This solves the technical problem of reducing the computational overhead of model retraining while supporting selective deletion of voice recordings. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the data forgetting method provided in the above embodiments, and will not be repeated here.

[0121] Furthermore, embodiments of this application also propose a computer program product, including a computer program, which, when executed by a processor, implements the steps of the data forgetting method described above.

[0122] The specific implementation of the computer program product in this application is basically the same as the various embodiments of the data forgetting method described above, and will not be repeated here.

[0123] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0124] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0125] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases, the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software sensor. This computer software sensor is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause an electronic device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of this application.

[0126] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A data forgetting method, characterized in that, The data forgetting method includes the following steps: In response to a data deletion request, identify the voice recordings to be deleted; The training slice associated with the voice record to be deleted is used as the target training slice, and the target sub-model trained by the target training slice is determined from the sub-models of the ensemble model. Remove the speech records to be deleted from the target training slice, and use the remaining data after deletion as the retraining sample set to retrain the target sub-model. In this case, the sub-models other than the target sub-model in the ensemble model remain unchanged. The retrained target sub-model is aggregated with the remaining sub-models to generate a new ensemble model after forgetting.

2. The data forgetting method as described in claim 1, characterized in that, After the step of determining the voice recording to be deleted, the method further includes: Obtain the record index of the voice recording to be deleted; The storage location of the original data is queried according to the record index, wherein the original data includes the audio file and / or text file corresponding to the voice record to be deleted in the storage layer; Delete the original data at the storage location.

3. The data forgetting method as described in claim 1, characterized in that, Before the step of using the training segment associated with the voice record to be deleted as the target training segment, the method further includes: Obtain the usage tag from the record index of the voice recording to be deleted; If the usage label indicates that the speech record to be deleted has been used in model training, the step of using the training slice associated with the speech record to be deleted as the target training slice is performed.

4. The data forgetting method as described in claim 1, characterized in that, The step of aggregating the retrained target sub-model with the unchanged sub-models to generate a new ensemble model after forgetting includes: Obtain the retrained target sub-model, and all sub-models in the ensemble model other than the target sub-model; The retrained target sub-model is combined with all other sub-models to form a new ensemble model after forgetting.

5. The data forgetting method as described in claim 1, characterized in that, Following the step of generating the new ensemble model after forgetting, the method further includes: Obtain the feature vector of the voice record to be deleted; The feature vector is input into the new ensemble model to obtain the prediction confidence of the new ensemble model for the voice record to be deleted; If the deviation between the prediction confidence level and the random guess level is less than or equal to a preset threshold, the forgetting is considered complete.

6. The data forgetting method as described in claim 1, characterized in that, The data forgetting method also includes: Get the retention period configured by the user; According to the retention period, the voice records are scanned, and the voice records whose generation time and retention period sum to the current time are identified as expired records; The expired records are used as voice records to be deleted, and the step of using the training slice associated with the voice records to be deleted as the target training slice is executed.

7. The data forgetting method as described in claim 1, characterized in that, Following the step of generating the new ensemble model after forgetting, the method further includes: Obtain the first prediction accuracy of the new ensemble model on a benchmark set, and the second prediction accuracy of the ensemble model on the benchmark set, wherein the benchmark set contains speech samples of multiple intent categories; If the difference between the first prediction accuracy and the second prediction accuracy is greater than a preset threshold, it is determined that the target sub-model corresponding to the target training slice has undergone feature shift during the retraining process. Trigger the compensation retraining of the target sub-model, wherein the compensation retraining uses samples in the benchmark set that belong to the same intent category as the speech record to be deleted as supplementary samples to perform incremental training on the target sub-model.

8. The data forgetting method as described in claim 1, characterized in that, After the step of using the training segment associated with the voice record to be deleted as the target training segment, the method further includes: Obtain the total number of times the data deletion request was initiated within the preset time window; If the total number of times is greater than a preset frequency threshold, the target training slices corresponding to each data deletion request are counted to obtain the number of times each training slice is hit. If the percentage of times a training slice is hit exceeds a preset threshold, a slice reconstruction instruction is generated, and retraining is performed on the sub-model corresponding to the training slice after data redistribution. The slice reconstruction instruction is used to redistribute the data in the first training slice to other training slices, where the first training slice is the training slice whose percentage of times it is hit exceeds the preset threshold.

9. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the data forgetting method as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the data forgetting method as described in any one of claims 1 to 8.

Images