Method and System for Index-Map-Based Selective Invalidation and Restoration Control of Large Language Model Parameters with Execution Attestation
The method addresses catastrophic collapse and unreliable verification in large language models by classifying parameter blocks and generating a target index map for precise control and indisputable verification, ensuring effective data removal and model integrity.
Patent Information
- Application Number
- KR1020260172283
- Authority / Receiving Office
- KR · KR
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2026-04-17
- Filing Date
- 2026-09-09
- Publication Date
- 2026-11-06
- Estimated Expiration
- 2046-09-09
AI Technical Summary
Conventional methods for removing specific data from large language models face issues such as catastrophic collapse, lack of quantitative criteria for restoration, and unreliable verification, leading to compromised model performance and integrity.
A method for selectively invalidating and restoring parameters in a large language model by classifying parameter blocks based on deletion and preservation contributions, applying different modifications, and generating a target parameter index map to ensure precise control and verification.
This approach prevents loss of preserved knowledge and retention of deleted knowledge, ensures consistent restoration, and provides indisputable verification through a digitally signed execution proof payload.
Smart Images

Figure 112026111858635-PAT00001_ABST
Abstract
Description
Technology Field
[0001] The present invention relates to a technology for controlling learning parameters of an artificial neural network, and more specifically, to a parameter control method and system that selectively invalidates activation paths of specific knowledge previously learned in a Large Language Model (LLM) at the parameter level while preserving the model's general language reasoning ability, locally restores reasoning paths damaged by invalidation, and outputs in a verifiable form that the entire process of invalidation and restoration was actually performed at specified coordinates. Background Technology
[0002] As the commercialization of large language models becomes widespread, issues are emerging regarding copyright-infringing data, sensitive personal information, and biased or harmful data included in the training process being reproduced in the models' responses. With the strengthening of the regulatory environment, including the right to erasure stipulated by the European Union's General Data Protection Regulation (GDPR) and the EU Artificial Intelligence Act (AI Act), model providers are facing technical obligations to retrospectively remove the influence of specific data.
[0003] As the commercialization of large language models becomes widespread, issues are emerging regarding copyright-infringing data, sensitive personal information, and biased or harmful data included in the training process being reproduced in the models' responses. With the strengthening of the regulatory environment, including the right to erasure stipulated by the European Union's General Data Protection Regulation (GDPR) and the EU Artificial Intelligence Act (AI Act), model providers are facing technical obligations to retrospectively remove the influence of specific data.
[0004] Machine unlearning technology is being researched to address this, but conventional technology has the following limitations.
[0005] First, suppressing or abstaining from the weights of specific parameters leads to a phenomenon known as Catastrophic Collapse, where the inference paths of normal knowledge associated with those parameters are also damaged, resulting in a decline in the model's overall linguistic ability. In particular, when the knowledge to be deleted and the knowledge to be preserved share the same parameter, a conflict arises where removing the parameter en masse causes the knowledge to be preserved to be lost, while preserving it causes the knowledge to be deleted to remain.
[0006] Second, conventional technology has not presented an implementation method that connects the techniques for searching target parameters and modifying weights into a single control flow, as these methods remain at the level of individual algorithms, including the process of restoring the damaged inference path after invalidation. Furthermore, the absence of quantitative criteria for determining whether to perform restoration leads to problems such as uniformly performing fine-tuning even when restoration is unnecessary, or failing to detect when restoration is required.
[0007] Third, there is a lack of means to objectively prove to regulatory authorities or data subjects that the data has actually been removed. Conventional verification methods are limited to behavioral-level verification, which involves observing the model's output to confirm that the knowledge subject to deletion is not reproduced. However, behavioral-level verification can yield varying results depending on the query method and fails to indicate exactly which parameters were changed and how within the model.
[0008] Furthermore, even if a method of recording and presenting parameter changes is adopted, there is a fundamental limitation in that there is no way to distinguish cases where the software generating the record does not actually perform the parameter change operations but merely produces and outputs a disguised record. In other words, the authenticity of the record depends on the good faith of the entity that generated it, making it difficult to use as reliable evidence in legal disputes.
[0009] Therefore, an integrated control means is required that can precisely invalidate and restore target knowledge at the parameter level without compromising the model's general reasoning ability, and output the fact of execution in an undisguised form. The problem to be solved
[0010] The first problem that the present invention aims to solve is to resolve the conflict arising from the sharing of parameters between knowledge to be deleted and knowledge to be preserved. To this end, the invention provides means for quantifying the degree to which each parameter block is involved in the reproduction of data to be deleted and the degree to which it is involved in general language inference, respectively, and for triadically classifying parameters according to the combination of the two to apply different modification methods for each classification.
[0011] The second problem that the present invention aims to solve is to maintain location information of parameters to be invalidated in a structured form that can be referenced by subsequent processing. Conventionally, the results of searching for target parameters were consumed only within the corresponding search step and could not be utilized for determining the restoration location or identifying the fact of change. The present invention generates an index map containing the coordinates of target parameters and allows each step of invalidation, restoration, and verification to refer to it in common, thereby enabling a single data structure to simultaneously define the target of control and the scope of verification.
[0012] The third problem that the present invention aims to solve is to quantitatively determine whether to perform a restoration step. To this end, the invention provides means for measuring a first indicator representing the degree of reproduction of data to be deleted and a second indicator representing the degree of preservation of general language reasoning ability for a model after invalidation, and comparing each indicator with a preset threshold value to branch to one of readjusting the invalidation range, performing restoration, or omitting restoration.
[0013] The fourth problem that the present invention aims to solve is to output, in an undisguised form, that a parameter change has actually been performed at a specified coordinate. To this end, the invention provides means for re-reading the changed parameters to verify them against a recorded value, performing the entire process within a trusted execution environment, and generating an execution proof payload that is digitally signed with a unique key dependent on the hardware of the environment and cannot be leaked externally. means of solving the problem
[0014] A parameter control method according to one aspect of the present invention for solving the above problem is a method for selectively invalidating and restoring parameters of a large language model performed by a computation server, comprising: (A) a knowledge coordinate localization step of generating a plurality of layer-specific activation maps by inputting data to be deleted and a preservation criterion input to the large language model, respectively, and calculating a deletion contribution and a preservation contribution by comparing the activation maps for the same parameter block, classifying the parameter block into one of a deletion neuron, a preservation neuron, and a collision neuron according to a combination of the result of comparing the deletion contribution and a first threshold value and the result of comparing the preservation contribution and a second threshold value, and generating a target parameter index map including a layer identifier, a channel identifier, a row / column or component position of a weight matrix, a classification tag, and the contribution corresponding to the deletion neuron and the collision neuron; (B) A weight suppression and target ablation step in which, by referring to the target parameter index map, at least some of the rows, columns, or components of the weight matrix corresponding to the deletion neuron are replaced with zero vectors, the components of the weight matrix corresponding to the collision neuron are attenuated by a damping coefficient proportional to the preservation contribution, and the weights corresponding to the preservation neuron are frozen, thereby blocking the activation path of knowledge related to the data to be deleted and generating a weight change log including the changed coordinates, the change method, the block hash before the change, and the block hash after the change; (C) A forgetting quality and utility verification step in which a first indicator and a second indicator are measured for the model after the target ablation step, and if the first indicator is less than a first threshold value, the ablation range is adjusted and the step (B) is re-executed, and if the first indicator is greater than or equal to the first threshold value and the second indicator is greater than or equal to the second threshold value, restoration is omitted, and if the second indicator is less than the second threshold value, it is determined that context collapse has occurred;(D) A knowledge inpainting execution step in which, when the above context collapse is determined, a low-rank adapter is inserted into the layer to which the coordinate specified by the target parameter index map belongs or an adjacent layer, and local fine-tuning is performed using the preserved data to form a knowledge inpainting bypass connection path, and a re-enactment history including the insertion position, rank, and adjustment range is generated; and (E) An execution proof step in which the parameters of the coordinates specified by the target parameter index map are re-read to compare the re-reading pattern and the record pattern bit by bit, and only if the comparison result matches, the deletion target data identifier, the commit value of the target parameter index map, the block hash before the change, the block hash after the change, the weight change log, the measurement values of the first and second indicators, and the re-enactment history are combined and then digitally signed with a unique key dependent on the hardware of the trusted execution environment to generate and output an execution proof payload, wherein steps (A) through (E) are performed within the trusted execution environment.
[0015] The above damping coefficient can be defined as the ratio of the conservation contribution to the sum of the deletion contribution and the conservation contribution of the collision neuron.
[0016] The above execution proof payload or its hash value may be recorded in a verifiable ledger, and the transaction identifier returned according to the record may be entered into the ledger record reference field of the above execution proof payload.
[0017] If the above comparison results are inconsistent, the above execution proof step may request the re-execution of the above weight suppression and target ablation step without performing an electronic signature.
[0018] The above-mentioned preservation data may be received from a preservation data provider connected to the above-mentioned computation server via a communication network, and the preservation data provider is located outside the above-mentioned trusted execution environment.
[0019] In addition, a system according to another aspect of the present invention for solving the above problem comprises a diagnostic module, a surgical module, a verification module, a rehabilitation module, a certification module configured to perform the above method, and an orchestrator that sequentially calls the same, wherein each module is encapsulated and operates within a trusted execution environment isolated from the main operating system. Effects of the invention
[0020] First, by triunely classifying parameter blocks based on a combination of deletion and preservation contributions and applying different transformation methods for each classification, it is possible to simultaneously avoid the loss of preserved knowledge due to batch removal and the retention of deleted knowledge due to batch preservation, even when the knowledge to be deleted and the knowledge to be preserved share parameters. In particular, by applying damping proportional to the preservation contribution to conflicting neurons, the magnitude of context collapse is suppressed by maintaining residual weights equal to the extent to which the corresponding parameter contributes to the preserved knowledge.
[0021] Second, since the target parameter index map simultaneously handles target designation for invalidation, determination of insertion locations for restoration adapters, and specification of the verification range for execution proofs, the three stages share the same coordinate system. Consequently, problems such as restoration being applied to locations unrelated to the invalidated area or the verification being inconsistent with the actual modification range are structurally prevented.
[0022] Third, by comparing the first and second indicators with threshold values and branching into readjustment of the resection range, performance of restoration, or omission of restoration, unnecessary fine-tuning is omitted to reduce computational resources, while selectively performing restoration only when context collapse occurs.
[0023] Fourth, by re-reading the changed parameters, comparing them bit by bit with the recorded values, and digitally signing the result with a hardware-dependent unique key, the spoofing of creating only records without actually performing parameter change operations is structurally blocked. Since the said unique key cannot be leaked outside the trusted execution environment, the very existence of the signature implies that verification has been passed within that environment.
[0024] Fifth, since the execution proof payload combines the deletion target identifier, modification coordinates, pre- and post-modification hashes, behavioral indicator measures, and execution signature into a single structure, it can present a single piece of evidence regarding what was targeted, at what location, what modification was performed, and what the result was. This is something that cannot be provided by conventional verification methods that rely solely on behavioral-level observation. Brief explanation of the drawing
[0025] FIG. 1 is a block diagram illustrating the overall configuration of a parameter control system according to one embodiment of the present invention. FIG. 2 is a flowchart illustrating the overall flow of a parameter control method according to one embodiment of the present invention. FIG. 3 is a detailed diagram illustrating the activation map comparison, contribution calculation, and three-way classification process according to one embodiment of the present invention. FIG. 4 is a structural diagram illustrating the data structure of a target parameter index map and its reference relationships according to an embodiment of the present invention. FIG. 5 is a detailed diagram illustrating the weight suppression and target resection process according to one embodiment of the present invention, contrasting the process before and after resection. FIG. 6 is a flowchart illustrating the process of verifying forgetting quality and utility and conditional restoration judgment according to one embodiment of the present invention. FIG. 7 is a detailed diagram illustrating the knowledge in-painting process in steps according to one embodiment of the present invention. FIG. 8 is a structural diagram illustrating the data structure and rereading comparison verification structure of an execution proof payload according to one embodiment of the present invention. FIG. 9 is a block diagram illustrating an execution structure inside a trusted execution environment and an external access blocking structure according to an embodiment of the present invention. FIG. 10 is a block diagram illustrating a multimodal expansion configuration according to one embodiment of the present invention. Specific details for implementing the invention
[0026] Preferred embodiments of the present invention will be described in detail below with reference to the attached drawings. The reference numerals assigned to components in each drawing refer to the same components throughout the specification.
[0027] 1. Overall System Configuration (Fig. 1)
[0028] Referring to FIG. 1, a machine forgetting integrated control system (100) according to one embodiment of the present invention is operated on a computing server (10) and includes a diagnostic module (110), a surgical module (120), a verification module (140), a rehabilitation module (130), a certification module (150), and an orchestrator (160). The entire system (100) is encapsulated and operates within a trusted execution environment (900) that is logically and physically isolated from the main operating system.
[0029] The above system (100) receives data to be deleted (Forgot Set) and retention criteria inputs. The data to be deleted is a set of data whose influence is to be removed from the model, and the retention criteria inputs are a set of criteria inputs to verify the general language reasoning ability that the model must maintain even after removal.
[0030] When the orchestrator (160) detects the input of the data to be deleted, it sequentially calls the diagnosis module (110), surgery module (120), verification module (140), rehabilitation module (130), and proof module (150). Accordingly, the entire process from the localization of knowledge coordinates to the generation of proof of execution can be performed without human intervention.
[0031] The data transmitted between each module is as follows. The diagnostic module (110) transmits the target parameter index map (700) to the surgery module (120), and the surgery module (120) transmits the post-resection model and weight change log (840) to the verification module (140). The verification module (140) transmits the context collapse determination signal to the rehabilitation module (130) and transmits the measurement values (851, 852) to the verification module (150). The rehabilitation module (130) transmits the rehabilitation history (860) to the verification module (150). The verification module (150) obtains an electronic signature from the hardware-dependent unique key (910), generates an execution proof payload (800), and outputs it to the verifiable ledger (950).
[0032] The preservation data providing unit (20) is an optional configuration located outside the trusted execution environment (900) and provides preservation data upon the request of the rehabilitation module (130). The preservation data is data that can be used to restore general language reasoning ability without including data to be deleted. However, the preservation data may be obtained from a dataset held by the computation server (10) itself, in which case the preservation data providing unit (20) may be omitted.
[0033] 2. Overall control flow (Fig. 2)
[0034] Referring to FIG. 2, the control flow of the present invention is initiated by the input of data to be deleted and branches according to conditions through a knowledge coordinate localization step (S210), a weight suppression and target ablation step (S220), and a forgetting quality and utility verification step (S250).
[0035] After the verification step (S250), a first judgment is performed. If the first index (FE) is less than the first reference value (Th_f), a resection re-execution loop is performed to return to the resection step (S220) after adjusting the resection range, assuming that the knowledge to be deleted has not been sufficiently invalidated.
[0036] If the first index is greater than or equal to the first threshold value, a second judgment is performed. If the second index (RU) is greater than or equal to the second threshold value (Th_r), it is because general language reasoning ability is maintained despite invalidation, so the restoration step is skipped and the process proceeds directly to the execution proof generation and output step (S260). This is called the rehabilitation skip path.
[0037] If the second indicator is less than the second threshold value, it is determined that context collapse has occurred, and the knowledge inpainting execution step (S240) is performed after the preservation data acquisition step (S230). After the inpainting is performed, a re-re-verification loop is performed to return to the verification step (S250), and during repetition, the range of the preservation data, the insertion position of the low-rank adapter, or the range of fine-tuning may be adjusted.
[0038] The entire process of steps S210 to S260 above is performed within the execution range of the trusted execution environment (900). However, the preservation data provision unit (20) is located outside the execution range, and only the request and the exchange of preservation data in step S230 pass through the boundary.
[0039] 3. Activation Map Comparison and Triangular Classification (Fig. 3)
[0040] Referring to FIG. 3, in the knowledge coordinate localization step (S210), the diagnostic module (110) inputs the data to be deleted into a large language model to generate a layer-by-layer activation map A_f, and inputs the retention criteria input into the same model to generate a layer-by-layer activation map A_r.
[0041] Next, the diagnostic module (110) compares the two activation maps for the same parameter block. Here, a parameter block refers to a set of parameters partitioned into units of a layer, channel, or component of a weight matrix. The diagnostic module (110) calculates a deletion contribution C_f from the activation intensity at A_f of the corresponding parameter block and calculates a preservation contribution C_r from the activation intensity at A_r. The deletion contribution indicates the degree to which the corresponding parameter block is involved in the reproduction of the data to be deleted, and the preservation contribution indicates the degree to which the corresponding parameter block is involved in general language inference.
[0042] The diagnostic module (110) compares the calculated contribution with the first threshold value (Th_1) and the second threshold value (Th_2), respectively, and classifies the parameter blocks as follows.
[0043] A parameter block in which C_f is greater than or equal to Th_1 and C_r is less than Th_2 is classified as a deletion neuron (310). This is a parameter block that exclusively engages with the data to be deleted and does not substantially engage with general language reasoning.
[0044] Parameter blocks where C_r is greater than or equal to Th_2 and C_f is less than Th_1 are classified as preservation neurons (320). These are parameter blocks that are necessary for general language reasoning and are substantially unrelated to the data to be deleted.
[0045] A parameter block in which C_f is greater than or equal to Th_1 and C_r is greater than or equal to Th_2 is classified as a conflicting neuron (330). This is a parameter block involved in both the data to be deleted and general language inference, and is the object that causes the conflict that the present invention seeks to resolve.
[0046] The diagnostic module (110) generates a target parameter index map (700) according to the classification result. The index map (700) includes coordinates corresponding to at least the deletion neuron (310) and the collision neuron (330), and may also include coordinates corresponding to the preservation neuron (320) as a freezing target.
[0047] The above thresholds Th_1 and Th_2 can be set according to the scale of the model, the nature of the data to be deleted, and the required level of forgetting, and can be adjusted together when adjusting the resection range according to the results of the verification step described later.
[0048] 4. Target Parameter Index Map (Fig. 4)
[0049] Referring to FIG. 4, the target parameter index map (700) includes a layer identification field (710), a channel identification field (720), a coordinate field (730), a classification tag field (740), and a contribution field (750).
[0050] The layer identification field (710) stores the number of the layer to which the target parameter belongs. The channel identification field (720) stores the channel number within that layer. The coordinate field (730) stores the row position, column position, or component position within the weight matrix. The classification tag field (740) stores whether the parameter block belongs to the classification of deletion, preservation, or collision. The contribution field (750) stores the deletion contribution C_f and the preservation contribution C_r calculated for the parameter block.
[0051] For example, a single record may be configured as follows: Layer 12, Channel 384, Row 7, Column 1024, Classification Tag Conflict, C_f=0.84, C_r=0.61. The above example is for illustrative purposes only and does not limit the scope of the invention.
[0052] The index map (700) is referenced by three subsequent processes. First, the surgery module (120) designates an ablation target by referencing the coordinate field (730) and classification tag field (740) of the index map (700). Second, the rehabilitation module (130) determines the insertion location of the low-rank adapter (520) by referencing specific coordinates of the index map (700). Third, the proof module (150) uses specific coordinates of the index map (700) as the hash target range for the execution proof payload (800).
[0053] In this way, a single data structure simultaneously defines the target of invalidation, the location of restoration, and the scope of proof, thereby ensuring that the three processes are performed in a consistent manner on the same coordinate system. In conventional technology, the results of the target search were consumed only within the search phase and could not be utilized to determine the location of restoration or to identify the fact of change; as a result, there were problems where restoration was applied to a location unrelated to the invalidation area or where the scope of proof did not match the actual change range.
[0054] 5. Weighted suppression and targeted resection (Fig. 5)
[0055] Referring to FIG. 5, in the weight suppression and target resection step (S220), the surgical module (120) applies different variations by classification by referring to the coordinate field (730) and classification tag field (740) of the target parameter index map (700).
[0056] At least some of the rows, columns, or components of the weight matrix corresponding to the deletion neuron (310) are replaced with zero vectors to form a zero vector replacement region (410). Accordingly, the activation path of knowledge related to the data to be deleted is blocked.
[0057] The components of the weight matrix corresponding to the collision neuron (330) are attenuated by a damping coefficient α that is proportional to the conservation contribution. That is, the weight W′ after damping is calculated as W′ = α·W, where the damping coefficient α is defined as α = C_r / (C_f + C_r).
[0058] According to the definition of the damping coefficient above, the greater the contribution of the corresponding collision neuron to general language reasoning, the closer the value of α is to 1, and the greater the residual weight is maintained; conversely, the greater the contribution to the data to be deleted, the closer the value of α is to 0, and the smaller the residual weight is maintained. Accordingly, the loss of preserved knowledge that occurs when collision neurons are removed in bulk and the retention of deleted knowledge that occurs when they are preserved in bulk can be simultaneously mitigated. The above damping can be selectively applied at the component level rather than to the entire weight connected to the collision neuron.
[0059] The weights corresponding to the preserved neurons (320) are frozen and are not changed during the above resection process.
[0060] The surgical module (120) generates a weight change log (840) during the resection process. The log (840) includes a list of resection coordinates, a resection method (distinction between zero vector permutation or attenuation), a block hash before resection (830), and a block hash after resection (831). The log (840) is used as a component of the execution proof payload (800) described later.
[0061] 6. Verification of forgetting quality and utility (Fig. 6)
[0062] Referring to FIG. 6, in the forgetting quality and utility verification step (S250), the verification module (140) measures two indicators in parallel for the model after ablation.
[0063] The first indicator (FE) is an indicator representing the degree to which the relevant knowledge is reproduced with respect to the data to be deleted and its modified query, and is calculated as a measurement value (851). The modified query refers to a query that indirectly induces the same knowledge without directly referring to the data to be deleted, and by measuring this together, it is possible to distinguish between superficial response blocking and actual knowledge invalidation.
[0064] The second indicator (RU) is an indicator representing the degree to which the model maintains response quality with respect to the preservation standard input, and is calculated as a measurement value (852).
[0065] The verification module (140) compares the first indicator with the first reference value (Th_f). If the first indicator is less than the first reference value, the ablation range expansion and adjustment is performed and the ablation step (S220) is re-executed. At this time, the adjustment may include at least one of lowering the threshold value Th_1, expanding the ablation target coordinates, or lowering the damping coefficient.
[0066] If the first indicator is greater than or equal to the first threshold value, the verification module (140) compares the second indicator with the second threshold value (Th_r). If the second indicator is greater than or equal to the second threshold value, rehabilitation is omitted and the process proceeds to the execution proof step (S260). If the second indicator is less than the second threshold value, it is determined to be context collapse and the process proceeds to the knowledge inpainting execution step (S240).
[0067] After performing knowledge inpainting, a rehabilitation-revalidation loop is performed to return to measurements for the post-resection model. During repetition, the range of preserved data, the insertion position of the low-rank adapter, or the range of fine-tuning may be adjusted.
[0068] By comparing the two aforementioned indicators with separate reference values and branching in three directions, cases where invalidation is insufficient and cases where invalidation is excessive can be distinguished, and different treatments can be applied accordingly. Conventionally, there was no quantitative criterion for determining whether to perform restoration, so fine-tuning was uniformly performed even when restoration was unnecessary; however, according to this embodiment, if the second indicator satisfies the criterion, restoration can be omitted, thereby reducing computational resources.
[0069] 7. Knowledge Inpainting (Fig. 7)
[0070] Referring to Fig. 7, the knowledge in-painting execution step (S240) unfolds into three phases: before rehabilitation, insertion of a low-rank adapter, and after rehabilitation.
[0071] In the model prior to rehabilitation, a defective parameter region (510) is formed due to resection, and the connection of the general context passing through that region is damaged.
[0072] The rehabilitation module (130) refers to the target parameter index map (700) and selects the layer to which the resection coordinates belong or an adjacent layer as the insertion location for the low-rank adapter (520). The low-rank adapter (520) is composed of the product of a d×r matrix and an r×k matrix, where r is set to a value smaller than d and k. Accordingly, the number of parameters to be adjusted is significantly reduced compared to the case where the entire original weight matrix is adjusted.
[0073] The rehabilitation module (130) performs local fine-tuning on the low-rank adapter (520) using the preserved data. Since the fine-tuning updates only the parameters of the inserted adapter without changing the original weights, the state of the zero-vector substitution region (410) formed by the resection is maintained.
[0074] In the model after rehabilitation, a knowledge inpainting bypass connection path (530) is formed. The path (530) is a connection that bypasses the missing parameter area (510) without directly restoring it to its original state, and accordingly, the activation path of the blocked deleted knowledge remains blocked while only the inference path of the general context is restored.
[0075] The rehabilitation module (130) generates a rehabilitation history (860) during the above process. The rehabilitation history (860) includes the insertion layer, rank, and adjustment range of the low-rank adapter (520) and is used as a component of the proof-of-execution payload (800).
[0076] 8. Proof of Execution Payload and Rereading Comparison Verification (Fig. 8)
[0077] Referring to FIG. 8, in the execution proof step (S260), the proof module (150) generates an execution proof payload (800) after performing a reread comparison verification. The proof module (150) is located inside the trusted execution environment (900).
[0078] The re-reading comparison verification is performed as follows. First, the proof module (150) re-reads the parameters of the coordinates specified by the target parameter index map (700). Then, the re-reading pattern P′ and the pattern P recorded in the ablation step are compared bit by bit.
[0079] If the comparison result matches, the proof module (150) proceeds with the signing procedure. If the comparison result does not match, the proof module (150) does not perform the signing and requests the re-execution of the excision step (S220). Accordingly, if the intended change at the specified coordinates is not actually made, the proof of execution is not generated.
[0080] The execution proof payload (800) includes a deletion target data identifier field (810), an index map commit field (820), a parameter block hash field before resection (830), a parameter block hash field after resection (831), a weight change log field (840), an indicator field (850), a rehabilitation history field (860), an execution signature field (870), and a ledger record reference field (880). Here, the indicator field (850) includes a first indicator measure (851) and a second indicator measure (852).
[0081] The index map commit field (820) stores the hash value for the entire target parameter index map (700). Accordingly, the set of coordinates covered by the proof cannot be changed retrospectively.
[0082] An electronic signature value based on a hardware-dependent unique key (910) is stored in the execution signature field (870). The unique key (910) is fused during the manufacturing stage and cannot be leaked outside the trusted execution environment (900). Therefore, the fact that the signature is valid means that the re-reading comparison verification has passed within the trusted execution environment.
[0083] According to this configuration, a disguise that creates and submits only a record without actually performing the parameter change operation is blocked. This is because the entity attempting the disguise cannot obtain the signature as it cannot pass the comparison verification, and cannot forge the signature as the unique key (910) cannot be leaked externally.
[0084] The generated execution proof payload (800) or its hash value may be recorded in a verifiable ledger (950). The transaction identifier returned according to the record is entered into the ledger record reference field (880). Accordingly, the time of creation of the execution proof cannot be retroactively determined.
[0085] The above execution proof payload (800) combines what was to be deleted (810), which coordinates were the target (820, 840), what was changed at those coordinates and how (830, 831), what the result of the change was (851, 852), where the restoration was applied (860), and whether the entire process was performed in an isolated environment (870) into a single structure.
[0086] 9. Internal structure of the Trusted Execution Environment (Fig. 9)
[0087] Referring to FIG. 9, an orchestrator (160), a diagnostic module (110), a surgical module (120), a verification module (140), a rehabilitation module (130), and a certification module (150) are sequentially arranged inside the trusted execution environment (900), and a hardware-dependent unique key (910) is arranged at the bottom.
[0088] Access to the inside of the trusted execution environment (900) from the main operating system and external processes is blocked at the hardware level. Accordingly, even if the main operating system is compromised or privileges are stolen, the integrity of the parameter control process is maintained, and attempts at impersonation from the outside are blocked.
[0089] Passing through the boundary of the above-mentioned trusted execution environment (900) is limited to the input of data to be deleted through a single input channel and the output of the execution proof payload (800) through a single output channel. Accordingly, intermediate data is prevented from leaking out or being injected from the outside.
[0090] A configuration in which the entire process of parameter control is performed within an isolated environment is significant in that it extends the scope of verification beyond the final output, the weight snapshot, to the execution process itself that generated the output. Since there may be multiple paths leading to the same weight state, it is not possible to determine what processing was performed based solely on the result; however, according to the present embodiment, the path itself is determined because the execution process is performed within an isolated environment and that fact is proven by a signature.
[0091] 10. Multimodal extension (Fig. 10)
[0092] Referring to FIG. 10, the present invention is not limited to text but can be extended to a multimodal neural network that processes images, audio, and video.
[0093] In a multimodal environment, the diagnostic module (110) generates an activation map corresponding to text tokens as well as image patches, voice frames, or video frames, and calculates deletion contribution and preservation contribution for the cross-modal attention weight area (610).
[0094] The surgery module (120) replaces or attenuates the target parameters included in the cross-modal attention weighting area (610) with a zero vector. The rehabilitation module (130) restores the multimodal context by inserting a low-rank adapter (520) into the visual inpainting area (620) or an adjacent layer of the area (610).
[0095] In this case, the target parameter index map (700) additionally includes a modal distinction field. The modal distinction field stores which modal the corresponding coordinates correspond to among text, image, audio, and video. Accordingly, a single index map can integrally manage targets across multiple modals.
[0096] The above embodiments are intended to aid in understanding the present invention, and the scope of the present invention is defined by the claims set forth below. A person skilled in the art will understand that various modifications and equivalent implementations are possible within the scope of the technical spirit of the present invention. Explanation of the symbols
[0097] [Systems and Hardware] 10: Computation Server 20: Preservation Data Provision Department 100: Machine Forgetting Integrated Control System 900: Trusted Execution Environment (TEE) 910: Hardware-dependent unique key 950: Verifiable Director [Module] 110: Diagnostic Module 120: Surgical Module 130: Rehabilitation Module 140: Verification Module 150: Proof Module 160: Orchestrator [Method Steps] S210: Knowledge coordinate localization step S220: Weighted suppression and targeted ablation steps S230: Data retention acquisition phase S240: Knowledge Inpainting Execution Step S250: Forgetfulness Quality and Utility Verification Steps S260: Proof of Execution Generation and Output Step [Internal elements of a neural network] 310: Delete Neuron 320: Preserving neuron 330: Collision neuron 410: Zero vector permutation region 510: Deficit parameter area 520: Low-Rank Adapter (LoRA Adapter) 530: Knowledge Inpainting Bypass Connection Path 610: Cross-modal attention weight area 620: Visual Inpainting Area [Target Parameter Index Map] 700: Target Parameter Index Map 710: Layer identification field 720: Channel identification field 730: Coordinate field 740: Classification tag field 750: Contribution Field [Execution Proof Payload] 800: Proof of Execution Payload 810: Data identifier field to be deleted 820: Index Map Commit Field 830: Pre-excision parameter block hash field 831: Parameter block hash field after excision 840: Weight change log field 850: Indicator field 851: First indicator measurement 852: Second indicator measurement 860: Rehabilitation History Field 870: Execution signature field 880: Ledger Record Reference Field
Claims
Claim 1 A method for selectively invalidating and restoring parameters of a large language model, performed by a computation server (10) operating a machine forgetting integrated control system (100), wherein (A) a diagnostic module (110) inputs data to be deleted and a preservation criterion input to the large language model, respectively, to generate a plurality of layer-specific activation maps, and compares the activation maps for the same parameter block to calculate a deletion contribution indicating the degree to which the parameter block is involved in the reproduction of the data to be deleted and a preservation contribution indicating the degree to which it is involved in general language inference, respectively, and classifies the parameter block into one of a deletion neuron (310), a preservation neuron (320), and a collision neuron (330) according to a combination of the result of comparing the deletion contribution with a first threshold and the result of comparing the preservation contribution with a second threshold, and a target parameter index including a layer identification field (710), a channel identification field (720), a coordinate field (730) storing the row, column, or component position of a weight matrix, a classification tag field (740), and a contribution field (750) storing the deletion contribution and the preservation contribution. (B) A knowledge coordinate localization step (S210) for generating a map (700); (B) A weight suppression and target resection step (S220) in which a surgical module (120), by referring to the coordinate field (730) and classification tag field (740) of the target parameter index map (700), substitutes at least some of the rows, columns, or components of the weight matrix corresponding to the deletion neuron (310) with a zero vector, attenuates the components of the weight matrix corresponding to the collision neuron (330) by a damping coefficient proportional to the preservation contribution, and blocks the activation path of knowledge related to the data to be deleted by freezing the weights corresponding to the preservation neuron (320), and generates a weight change log (840) including a resection coordinate list, a resection method, a parameter block hash before resection, and a parameter block hash after resection;(C) A verification module (140) measures a first indicator representing the degree of reproduction of knowledge regarding the data to be deleted and its modified query, and a second indicator representing the degree of maintenance of response quality for the preservation standard input, for the model after the target ablation step (S220); if the first indicator is less than the first standard value, the ablation range is adjusted and the (B) step is re-executed; if the first indicator is greater than or equal to the first standard value and the second indicator is greater than or equal to the second standard value, the (D) step described later is omitted; and if the first indicator is greater than or equal to the first standard value and the second indicator is less than the second standard value, it is determined that context collapse has occurred (S250); (D) If context collapse is determined, a rehabilitation module (130) selects the layer to which the coordinates specified by the target parameter index map (700) belong or an adjacent layer as the insertion location to insert a low-rank adapter (520), and uses the preservation data to [include] in the low-rank adapter (520 A knowledge inpainting execution step (S240) that performs local fine-tuning, wherein the local fine-tuning forms a knowledge inpainting bypass connection path (530) while maintaining the state of the zero vector substitution region (410) formed in step (B) by updating only the parameters of the low-rank adapter (520) without changing the original weights of the large language model, generates a rehabilitation history (860) including the insertion position, rank, and adjustment range, and returns to step (C) to remeasure the first indicator and the second indicator;and (E) for a model after it is determined in step (C) that the first indicator is greater than or equal to the first reference value and the second indicator is greater than or equal to the second reference value, the proof module (150) re-reads the parameters of the coordinates specified by the target parameter index map (700) and compares the re-reading pattern with the pattern recorded in step (B) bit by bit, and only if the comparison result matches, combines the deletion target data identifier (810), the index map commit (820) which is the hash value for the target parameter index map (700), the parameter block hash before resection (830), the parameter block hash after resection (831), the weight change log (840), the measurement value of the first indicator (851) and the measurement value of the second indicator (852), and the rehabilitation history (860), and then electronically signs with a unique key (910) that is dependent on the hardware of the trusted execution environment (900) and does not leak to the outside, thereby providing an execution proof including an execution signature (870). A method for selectively invalidating and restoring parameters of a large language model, comprising: an execution proof step (S260) that generates and outputs a payload (800), and requests the re-execution of step (B) without performing the electronic signature if the comparison result is inconsistent, wherein steps (A) through (E) are performed within the trusted execution environment (900). Claim 2 A method for selectively invalidating and restoring parameters of a large language model, characterized in that, in claim 1, the attenuation coefficient α of step (B) is defined as the ratio of the preservation contribution to the sum of the deletion contribution and the preservation contribution of the corresponding collision neuron (330), and the weight after attenuation is calculated by multiplying the weight before attenuation by the attenuation coefficient α. Claim 3 A method for selectively invalidating and restoring parameters of a large language model, wherein, in claim 1, the classification of step (A) is characterized by classifying a parameter block in which the deletion contribution is greater than or equal to the first threshold and the preservation contribution is less than the second threshold as a deletion neuron (310), classifying a parameter block in which the preservation contribution is greater than or equal to the second threshold and the deletion contribution is less than the first threshold as a preservation neuron (320), and classifying a parameter block in which the deletion contribution is greater than or equal to the first threshold and the preservation contribution is greater than or equal to the second threshold as a collision neuron (330). Claim 4 delete Claim 5 A method for selectively invalidating and restoring parameters of a large language model, wherein, in claim 1, the adjustment of the ablation range of step (C) comprises at least one of the downward reduction of the first threshold value, the expansion of the ablation target coordinates included in the coordinate field (730), and the downward reduction of the damping coefficient. Claim 6 A method for selectively invalidating and restoring parameters of a large language model according to claim 1, characterized in that if the second indicator remeasured in step (C) after step (D) is less than the second reference value, at least one of the range of the preserved data, the insertion position of the low-rank adapter (520), and the range of the local fine-tuning is adjusted and step (D) is repeated. Claim 7 A method for selectively invalidating and restoring parameters of a large language model, characterized in that, in claim 1, the method further includes the step of recording the execution proof payload (800) or its hash value generated in step (E) in a verifiable ledger (950), and writing the transaction identifier returned according to the record in the ledger record reference field (880) of the execution proof payload (800). Claim 8 A method for selectively invalidating and restoring parameters of a large language model, characterized in that, in claim 1, the preservation data is received from a preservation data provider (20) located outside the trusted execution environment (900) and connected to the computation server (10) via a communication network, and the data passing through the boundary of the trusted execution environment (900) is limited to the input of the data to be deleted, the exchange of the preservation data, and the output of the execution proof payload (800). Claim 9 A parameter selective invalidation and restoration control system (100) for a large language model, comprising a trusted execution environment (900) that is logically and physically isolated from the main operating system and has external access blocked; and includes a diagnostic module (110), a surgical module (120), a verification module (140), a rehabilitation module (130), a proof module (150) encapsulated and operating within the above-mentioned trusted execution environment (900), and an orchestrator (160) that sequentially calls each of the above modules; wherein the diagnostic module (110) calculates deletion contribution and preservation contribution through a comparison of layer-by-layer activation maps for data to be deleted and preservation criteria inputs, classifies parameter blocks into deletion neurons (310), preservation neurons (320), and collision neurons (330) according to the combination of the comparison results of each contribution and a first threshold and a second threshold, and is configured to generate a target parameter index map (700) including a layer identification field (710), a channel identification field (720), a coordinate field (730), a classification tag field (740), and a contribution field (750); and the surgical module (120) refers to the target parameter index map (700) to the deletion neuron (310) It is configured to form a zero vector substitution region (410) by substituting the corresponding weights with zero vectors, attenuate the weights corresponding to the collision neuron (330) with a damping coefficient proportional to the conservation contribution, freeze the weights corresponding to the conservation neuron (320), and generate a weight change log (840); the verification module (140) measures the first indicator and the second indicator and branches into one of adjusting the resection range, omitting rehabilitation, or determining context collapse, and is configured to re-measure the first indicator and the second indicator after local fine-tuning by the rehabilitation module (130); the rehabilitation module (130) inserts a low-rank adapter (520) into the layer or adjacent layer to which the target parameter index map (700) belongs, andA system for selective invalidation and restoration of parameters of a large language model (100), characterized in that it is configured to maintain the state of the zero vector substitution region (410) and generate a rehabilitation history (860) by performing local fine-tuning that updates only the parameters of the low-rank adapter (520) without changing the original weights of the large language model, and the proof module (150) is configured to re-read the parameters of the coordinates specified by the target parameter index map (700) for the model after the verification module (140) determines that the first indicator is greater than or equal to the first reference value and the second indicator is greater than or equal to the second reference value, compare them with the record pattern in bit units, generate an execution proof payload (800) electronically signed with a unique key (910) dependent on the hardware of the trusted execution environment (900) only when the comparison result matches, and if the comparison result does not match, it is configured not to perform the electronic signature and to request the surgery module (120) to re-perform the resection.
Citation Information
Patent Citations
Medical data encryption system for multi-modal large language model
CN117951746A
Text generation method and device, electronic equipment and medium
CN120087336A
Channel sensitivity driven large language model privacy memory forgetting method
CN121637564A
Lightweight deployment method and system for large edge model
CN121683899A
Learning apparatus and method
US20240028902A1
