Theft notification system for a generative machine learning module, method for theft detection, and method for configuring a detector
Patent Information
- Application Number
- PCT/EP2026/053353
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-02-24
- Filing Date
- 2026-02-09
- Publication Date
- 2026-08-27
Smart Images
Figure EP2026053353_27082026_PF_FP_ABST
Abstract
Description
[0001] 202501217
[0002] 1
[0003] Description
[0004] Theft detection system for a generative machine learning module, methods for theft detection and methods for detector configuration
[0005] Complex machines, such as robots, motors, construction systems, manufacturing plants, machine tools, logistics systems, turbines, energy supply systems, or motor vehicles, typically require complex control, design, operating, and / or monitoring procedures for productive and intelligent operation. For this purpose, machine learning techniques, especially so-called generative AI models (AI: Artificial Intelligence), are increasingly being used.
[0006] Such generative AI models can be used in generative machine learning modules to generate optimized control actions for controlling machines, optimized design data for manufacturing products, or to implement intelligent user interfaces. Generative AI models, such as large language models, are often pre-trained with considerable effort and provided in this form. The pre-trained AI models can then be adapted to specific tasks using various, generally less complex, methods.
[0007] Training such generative AI models typically requires very large amounts of training data, significant computing resources, and a great deal of specific expert knowledge. Therefore, there is a strong interest in protecting trained AI models or the training information they contain against uncontrolled or unauthorized distribution or use, and / or in detecting theft.
[0008] It is known that, for theft detection of neural networks, their neural weights are marked with a unique digital watermark before they are released. Based on this watermark, a neural network can then be checked to determine whether it originates from the user of the watermark or not.
[0009] However, such watermarking is often ineffective for generative AI models, especially pre-trained models. This is particularly true because the internal structure of generative AI models is frequently not accessible in detail. Furthermore, many generative AI models are not deterministic, meaning that different responses can be generated from the same input.
[0010] The object of the present invention is to provide a theft detection system for a generative machine learning module, a method for theft detection, and a method for configuring a detector for recognizing a generative machine learning module, which allow for more effective protection of generative machine learning modules.
[0011] This problem is solved by a method with the features of claim 1, by a method with the features of claim 12, by a theft detection system with the features of claim 13, by a computer program product with the features of claim 14, and by a computer-readable storage medium with the features of claim 15.
[0012] To configure a detector designed to recognize a first generative machine learning module, a learning-based data sequence generator produces a large number of training data sequences. These sequences are then fed into the first generative machine learning module and into a separate, second generative machine learning module. The first generative machine learning module generates initial response data sets, and the second generative machine learning module generates subsequent response data sets. The first and / or the second generative machine learning module can implement or include, in particular, a large language model (LLM), a generative pre-trained transformer (GPT), a generative adversarial network (GAN), and / or a diffusion model.The first and / or second response data sets can, in particular, represent response texts, time series, geometries, audio sequences, speech outputs, images, and / or video sequences. According to the invention, the variance of a first statistical distribution of the first response data sets, as well as the overlap of a second statistical distribution of the second response data sets with the first statistical distribution, are determined, and the data sequence generator is trained to reduce the variance and the overlap. Furthermore, a test data sequence generated by the trained data sequence generator is assigned to a response data set generated from it by the first generative machine learning module.The detector is then configured to feed the test data sequence into a generative machine learning module under test, compare a resulting response data set with the response data set assigned to the test data sequence, and output a detection signal accordingly. 202501217.
[0013] 3
[0014] To detect theft from a generative machine learning module, a detector is configured according to the procedure described above. The configured detector then feeds the test data sequence into the generative machine learning module under test and reads in a resulting response data set. Furthermore, a deviation between the read response data set and the response data set assigned to the test data sequence is determined. Depending on the determined deviation, a detection signal is then output.
[0015] To carry out the methods according to the invention, a theft alarm system, a computer program product and a computer-readable, preferably non-volatile storage medium are provided.
[0016] The methods and the theft detection system according to the invention can each be executed or implemented using one or more processors. In particular, application-specific integrated circuits (ASICs), digital signal processors (DSPs), and / or so-called "field-programmable gate arrays" can be used. Furthermore, the methods according to the invention can be executed, at least partially, in a cloud and / or in an edge computing environment.
[0017] A particular advantage of the invention is that it allows the protection of generative machine learning modules that do not operate deterministically and / or whose internal structure is inaccessible. In particular, generative machine learning modules trained or pre-trained by third parties can be effectively protected against uncontrolled distribution. Extensive retraining is generally not required. Furthermore, the invention is flexibly applicable and, in particular, not limited to artificial neural networks.
[0018] The response data set associated with the test data sequence can be considered a watermark, the output of which is induced by the test data sequence. If the training of the data sequence generator is aimed at reducing variance, this watermark can often be reproduced with sufficient reliability by the first generative machine learning module. Furthermore, since the training of the data sequence generator is also aimed at reducing overlap, it can often be assumed that the response data sets derived from the test data sequence by other generative machine learning modules will differ sufficiently from this watermark.
[0019] 4
[0020] Advantageous embodiments and further developments of the invention are specified in the dependent claims.
[0021] According to an advantageous embodiment of the invention, a weighted sum of the dispersion and the overlap can be formed, and the training of the data sequence generator can be directed to minimize the weighted sum.
[0022] Training can be stopped as soon as the weighted sum falls below a predefined threshold. The weighted sum can be calculated using a cost function, which is then used to train the data sequence generator.
[0023] To carry out such training given a cost function, a variety of efficient learning methods are available, in particular supervised learning.
[0024] Complementary to a cost function to be minimized, a reward function to be maximized can also be used for training in an equivalent manner.
[0025] According to a further advantageous embodiment of the invention, the determined variance can be compared with a predetermined first threshold and the determined overlap with a predetermined second threshold. The training can be terminated as soon as both the first and the second thresholds are undershot.
[0026] The first threshold can be used to define a minimum volatility or stability level for the initial response data sets. This threshold can be chosen to ensure that the resulting initial response data sets correspond sufficiently closely to the response data set assigned to the test data sequence and can therefore be reliably identified. Conversely, the second threshold can be used to define a minimum probability of confusion between the second response data sets and the initial response data sets. This second threshold can be chosen to ensure that the probability of confusion is no greater than a predefined error probability of the detector.
[0027] Advantageously, the trained data sequence generator can generate multiple, especially different, test data sequences. Accordingly, 202501217
[0028] 5
[0029] Each generated test data sequence is assigned a corresponding response data set generated by the first generative machine learning module. The detector can then be configured to feed the test data sequences into a generative machine learning module under test, compare each resulting response data set with the response data set assigned to the respective test data sequence, and output the detection signal accordingly.
[0030] By using multiple test data sequences to test a generative machine learning module, the error probability of the detector can often be significantly reduced.
[0031] Furthermore, individual training data sequences can be fed into the first and second generative machine learning modules multiple times. Accordingly, the first statistical distribution can be calculated based on the first response data sets of the repeatedly fed training data sets, and the second statistical distribution can be calculated based on the second response data sets of the repeatedly fed training data sets.
[0032] By repeatedly feeding the same training data sequence into the first and second machine learning modules and evaluating the resulting response datasets, the indeterminacy of these modules can be estimated. In particular, the variance can be used as a measure of the indeterminacy of the first machine learning module. In this way, the detection probability and / or error probability of the detector can be estimated and specifically reduced by training the data sequence generator.
[0033] According to a further advantageous embodiment of the invention, the first response data sets can each be encoded by a first vector and the second response data sets by a second vector. Accordingly, the variance can be derived from a statistical variance of the first vectors and the overlap from distances between the first vectors and the second vectors.
[0034] The encoding of response data sets or input data sequences of generative machine learning modules using vectors is often referred to as embedding, vectorization, or simply embedding. In large language models, word sequences formulated in natural language are often read in and / or output as input data sequences and / or response data sets. In such cases, known methods can be used.
[0035] 6
[0036] For word embedding, methods such as Word2Vec, Continuous Bag of Words (CBOW), or Skip-Gram can be used. For other generative machine learning models, a variety of other efficient vectorization methods are available, for example, for encoding time series, geometries, images, audio sequences, or video sequences.
[0037] Using the first vectors, the variance can be calculated using standard statistical methods, for example, by determining the Euclidean distances of the first vectors from their mean. The overlap can also be determined using standard statistical methods based on the first and second vectors, for example, by calculating normalized scalar products between the first and second vectors.
[0038] In particular, the overlap can be derived from a Szymkiewicz-Simpson coefficient and / or a Jaccard coefficient of the first and second statistical distributions.
[0039] Furthermore, the overlap can be derived from a product of the first statistical distribution with the second statistical distribution.
[0040] According to a further advantageous embodiment of the invention, the data sequence generator can be trained using a reinforcement learning method.
[0041] In particular, by means of a so-called model-free learning method, such as a Monte Carlo method, a SARSA method (SARSA: State-Action-Reward-State-Action) or a so-called Q-learning method, or by means of a rollout-based learning method based on state prediction.
[0042] According to a further advantageous embodiment of the invention, several second generative machine learning modules can be provided, into each of which the plurality of training data sets are fed, and for each of which an overlap of a respective second statistical distribution of the second response data sets with the first statistical distribution is determined. Accordingly, the data sequence generator can be trained to reduce the variance and the overlaps.
[0043] In particular, a maximum overlap can be selected from the overlaps, and the data sequence generator can be trained to reduce the variance as well as the maximum overlap. 202501217
[0044] 7
[0045] Alternatively or additionally, the overlaps can be added to a total overlap, and the data sequence generator can be trained to reduce the variance as well as the total overlap.
[0046] By reducing the overlaps with multiple, especially many, second generative machine learning modules, the probability of confusion with the first generative machine learning module can often be significantly reduced.
[0047] An embodiment of the invention is explained in more detail below with reference to the drawing. The drawings illustrate each embodiment schematically.
[0048] Figure 1 shows the configuration of a detector for detecting a generative machine learning module.
[0049] Figure 2: A theft detection system for a generative machine learning module and
[0050] Figure 3 shows another embodiment of a theft detection system for a generative machine learning module.
[0051] Insofar as the figures use the same or corresponding reference symbols, these reference symbols denote the same or corresponding entities, which may be described, implemented or designed in particular as in connection with the figure in question.
[0052] Figure 1 illustrates the configuration of a detector DET for detecting a first generative machine learning module LLM1. The detector DET has one or more processors PROC for executing process steps according to the invention and one or more memory MEMs for storing the data to be processed by the detector DET.
[0053] For the present embodiment, it is assumed that the first generative machine learning module, LLM1, comprises or implements a large language model, such as GPT-4, GPT-3, BLOOM, LLaMA, T5-11B, PaLM-E, DeepSeek V3, DeepSeek R1, or Gemini Pro. The first generative machine learning module, LLM1, is trained or pre-trained to generate related response texts from input texts. The input texts are fed into the large language model in the form of data sequences.
[0054] 8
[0055] Input texts are fed in, and the resulting response texts are output in the form of response data records. The input texts and / or the output texts can be formulated in one or more natural languages.
[0056] Such large language models are often abbreviated as "LLM" (Large Language Model) and have been generally available or usable for some time. In particular, a so-called generative pre-trained transformer, abbreviated GPT (Generative pre-trained Transformer), can be used as a large language model.
[0057] Alternatively or additionally, the first generative machine learning module LLM1 can implement or include a generative adversarial network (GAN), a diffusion model, or models with other architectures.
[0058] The input data sequences and / or the response data sets can be used to encode not only natural language text, but also time series, geometries, audio sequences, speech outputs, images or video sequences.
[0059] The response data sets can be used, in particular, to generate optimized control actions that operate a machine, or optimized design data that prompts a manufacturing plant to produce an optimized product. Furthermore, the response data sets can be used to control an intelligent user interface.
[0060] Since training a generative machine learning module, here LLM1, usually requires considerable computing resources, there is a great interest in protecting a trained generative machine learning module against uncontrolled or unauthorized distribution or use and / or in detecting theft.
[0061] In this context, the DET detector should be configured to reliably detect the existing first generative machine learning module LLM1 and to distinguish it from other generative machine learning modules. A theft detection system can then be implemented using the configured DET detector.
[0062] For this purpose, the response behavior of the first generative machine learning module LLM1 will be compared with the response behavior of one or more other generative machine learning modules. A recognition reliability of 202501217 is expected.
[0063] 9
[0064] The discrimination capability of the configured detector DET generally increases as more other generative machine learning modules are included in this comparison.
[0065] Figure 1 shows three examples of such generative machine learning modules, which will be referred to below as second generative machine learning modules (LLM2). It is assumed, for the sake of argument, that the second generative machine learning modules (LLM2) also include or implement a large language model.
[0066] When comparing the response behavior of large language models, it can be observed that different large language models often give the same or similar answers to the same input data sequence. For example, this applies to questions about generally known facts, such as the year of the first manned moon landing.
[0067] To achieve high discrimination, input data sequences should be found where the responses of the second generative machine learning module (LLM2) differ significantly from the responses of the first generative machine learning module (LLM1). Simultaneously, the responses of the first generative machine learning module (LLM1) should be reproducible.
[0068] For this purpose, a learning-based data sequence generator SGEN of the detector DET is trained to generate such input data sequences.
[0069] In this context, training is generally understood as the optimization of a behavior or function of a learning-based system, in this case, the data sequence generator SGEN. The behavior or function is determined by adjustable parameters of the learning-based system. During training, these parameters are adjusted or optimized so that the behavior or function of the learning-based system fulfills one or more predefined criteria as effectively as possible. For example, in learning-based control systems, the success of the generated control actions could be a criterion, or in prediction systems, the accuracy of the predictions. Through training, for instance, the network structures of neurons in a neural network and / or the weights of connections between neurons can be adjusted or optimized to best meet the predefined criteria. Training can thus be understood as an optimization problem.
[0070] 10
[0071] For such optimization problems in the field of machine learning, a variety of efficient optimization methods are available, in particular gradient-based optimization methods, gradient-free optimization methods, backpropagation methods, particle swarm optimizations, genetic optimization methods and / or population-based optimization methods.
[0072] In this case, the data sequence generator SGEN is configured to generate a large number of data sequences for input into the first generative machine learning module LLM1. The generation method used by the data sequence generator SGEN can be varied or configured using a variety of generation parameters. The data sequences are preferably generated in the form of a representational vector.
[0073] For the present embodiment, it is assumed that the data sequence generator SGEN comprises an artificial neural network whose parameters are optimized using a reinforcement learning method. Alternatively or additionally, the data sequence generator SGEN can comprise a recurrent neural network, a convolutional neural network, a perceptron, a Bayesian neural network, an autoencoder, a variational autoencoder, a Gaussian process, a deep learning architecture, a support vector machine, a data-driven regression model, a physical model, and / or a decision tree.
[0074] At the start of the training of the data sequence generator SGEN, its generation parameters can be initialized with random values or with predefined values. These predefined values can, in particular, be derived from a previous operation of the data sequence generator SGEN.
[0075] During training, the data sequence generator SGEN generates a multitude of different training data sequences TS, depending on its generation parameters. Each generated training data sequence TS is then fed N times as input data into both the first generative machine learning module LLM1 and all second generative machine learning modules LLM2.
[0076] By repeating a training data sequence TS N times, a statistical distribution of its response behavior can be estimated, particularly for non-deterministic generative machine learning modules. The number N of repetitions gives 202501217
[0077] 11
[0078] In this sense, a sample size is specified and can, for example, be set to 10, 100, or 1000. Preferably, N is chosen depending on the expected indeterminacy and / or a measured variance of the response behavior of the first and / or the second generative machine learning modules LLM1 and LLM2, respectively. Preferably such that statistically representative distributions of the response behaviors can be determined.
[0079] From the N repetitions of a respective training data sequence TS, the first generative machine learning module LLM1 derives N response data sets A1, and the second generative machine learning module LLM2 derives N response data sets A2. The response data sets A1 and A2 are each encoded by a representational vector.
[0080] The response datasets A1 and A2 are each transmitted to a discriminator DC of the detector DET. The discriminator DC determines a variance VAR of a statistical distribution SD1 from the N response datasets A1. The variance VAR can be calculated, for example, using standard statistical methods from the Euclidean distances of the representing vectors of the response datasets A1 from their mean. Furthermore, the discriminator DC determines a machine learning module-specific overlap OL of a statistical distribution SD2 of each of the N response datasets A2 with the statistical distribution SD1 for each of the second generative machine learning modules LLM2.The respective overlap OL can also be determined using standard statistical methods, for example, by means of normalized scalar products between the representing vectors of the response data sets A1 and A2 and / or by means of a product of the statistical distribution SD1 with the respective statistical distribution SD2. Preferably, the respective overlap OL can be derived from a Szymkiewicz-Simpson coefficient and / or a Jaccard coefficient of the statistical distributions SD1 and SD2.
[0081] The overlaps OL of the various second generative machine learning modules LLM2 can each be understood as the probability of confusion between the response behavior of a respective second generative machine learning module LLM2 and the response behavior of the first generative machine learning module LLM1.
[0082] In the present embodiment, the overlaps OL of the various second generative machine learning modules LLM2 are compared by the discriminator DC. A maximum overlap OLM is determined, i.e., the overlap OL with the largest overlap value. The maximum overlap represents, in a sense, a 202501217
[0083] 12
[0084] This represents the maximum probability of confusion between the first generative machine learning module LLM1 and the second generative machine learning modules LLM2.
[0085] Alternatively or additionally, the overlaps OL from the discriminator DC can also be added to obtain a total overlap. The latter can also be interpreted as a measure of the probability of confusion between the first generative machine learning module LLM1 and the second generative machine learning module LLM2.
[0086] The maximum overlap OLM and the variance VAR are used to calculate a weighted sum SC using the discriminator DC, for example, SC = W*VAR + (1-W)*OLM. Here, W represents a weighting factor of, for example, 1 / 2. Instead of the maximum overlap OLM, the total overlap can also be used to calculate the weighted sum SC.
[0087] Apparently, the weighted sum SC reflects both a dispersion of the response behavior of the first generative machine learning module LLM1 and a probability of confusion with the other generative machine learning modules LLM2.
[0088] The weighted sum SC is calculated for the N repetitions of a training data sequence TS. Accordingly, further weighted sums SC are calculated for the multitude of different training data sequences TS as described above.
[0089] To train the data sequence generator SGEN, the weighted sums SC from the discriminator DC are fed back to the data sequence generator SGEN, as indicated by a dashed arrow in Figure 1. During training, the generation parameters, or neural weights, of the data sequence generator SGEN are adjusted such that the weighted sums SC are minimized, at least on average. A variety of efficient learning methods, particularly reinforcement learning methods, can be used for this purpose.
[0090] The training can be terminated as soon as the variance VAR falls below a first predefined threshold, at least on average, and the maximum overlap OLM falls below a second predefined threshold, at least on average. The first threshold can be chosen so that the response data sets A1 are sufficiently reproducible and can therefore be reliably identified. Conversely, the second threshold can define a minimum probability of confusion between the 202501217.
[0091] 13
[0092] Response data sets A1 and A2 are specified. In particular, the second threshold can be chosen such that the probability of confusion is no greater than an acceptable error probability of the detector DET.
[0093] Alternatively, the training can be terminated as soon as the weighted sum SC falls below a predetermined threshold, at least on average.
[0094] The variance VAR determined in the trained state of the data sequence generator SGEN, the determined maximum overlap OLM and, if applicable, the individual overlaps OL are stored in the detector DET for later testing of generative machine learning modules.
[0095] Through training, the data sequence generator SGEN is typically enabled to generate data sequences where the responses of the first generative machine learning module, LLM1, are largely reproducible, but differ significantly from the responses of the second generative machine learning module, LLM2. The responses of the first generative machine learning module, LLM1, can thus be considered its characteristic watermark.
[0096] The trained data sequence generator SGEN then preferably generates several data sequences as test data sequences CS, each of which is fed into the first generative machine learning module LLM1 as input data. From each test data sequence CS, the first generative machine learning module LLM1 derives a corresponding response data set CA and assigns it to the respective test data sequence CS.
[0097] The resulting pairs (CS, CA) from a respective test data sequence CS and the respective associated response data set CA can be used by the detector DET to check whether a given generative machine learning module matches the first generative machine learning module LLM1 or not.
[0098] For this purpose, the detector DET is configured to feed the test data sequences CS into each generative machine learning module under test, compare each induced response data set with the corresponding response data set CA, and output a detection signal accordingly. The detection signal indicates that the generative machine learning module under test has been recognized as the first generative machine learning module LLM1. The detection signal can, in particular, be used to...
[0099] 14
[0100] This will be output if, for all test data sequences CS, an induced response data record deviates only slightly from the respective assigned response data record CA. It is expected that the probability of error will decrease as more test data sequences are used.
[0101] Figure 2 illustrates a first implementation variant of a corresponding theft detection system (DMS) for a generative machine learning module, here LLM1. The theft detection system (DMS) is configured to report a generative machine learning module that matches the first generative machine learning module, LLM1. If the latter has been placed on the market without the consent of the rightful owner, a theft can thus be detected.
[0102] The theft detection system DMS includes the detector DET, which, as described above, was configured to recognize the first generative machine learning module LLM1. According to the first implementation variant, the generated pairs (CS, CA) from a respective test data sequence CS and the corresponding response data set CA were stored in the detector DET.
[0103] It is assumed that a generative machine learning module (LLM) to be tested is linked to the theft reporting system (DMS).
[0104] To test the generative machine learning module LLM, the stored test data sequences CS are fed into the generative machine learning module LLM as input data sequences by the detector DET. As a result, the generative machine learning module LLM derives a corresponding response data set A from each test data sequence CS, which is then transmitted to the detector DET.
[0105] The DET detector compares each response data set A with the response data set CA assigned to the respective test data sequence CS. This determines the respective deviation D between response data sets A and CA. The deviation D can be derived, for example, from the Euclidean distance between the representing vectors of the respective response data sets A and CA, e.g., according to D = (A - CA). 2 or D=|A-CA|.
[0106] The determined deviations D are fed into an evaluation module EV of the detector DET. The evaluation module EV checks the fed-in deviations D by 202501217
[0107] 15
[0108] Comparison with the variance (VAR) stored after training, and, if applicable, depending on the stored maximum overlap (OLM) and / or the stored overlaps (OL). If the deviations (D) lie within the range quantified by the stored variance (VAR) in a statistically significant manner, and / or occur significantly more frequently within this range than would be expected based on the stored maximum overlap (OLM) or the overlaps (OL), the evaluation module (EV) generates a detection signal (DS). Otherwise, a complementary signal is generated. The detection signal (DS) or the complementary signal is output by the theft alarm system (DMS).
[0109] The detection signal DS indicates that the generative machine learning module LLM under test is highly likely to match the first generative machine learning module LLM1.
[0110] The emitted detection signal DS can, in particular, trigger an alarm signal via an alarm transmitter. The alarm signal can, for example, be transmitted to the creator of the first generative machine learning module LLM1 to inform the relevant department about the detection of a matching machine learning module.
[0111] Alternatively or additionally, as a result of the output of the detection signal DS, a function of a machine controlled by the generative machine learning module LLM under test, or of another technical system controlled by it, may be restricted or suppressed.
[0112] Figure 3 illustrates a second embodiment of a theft detection system (DMS) according to the invention. Unlike the theft detection system described in connection with Figure 2, in the second embodiment the pairs (CS, CA) from a respective test data sequence CS and the respective associated response data set CA are generated as needed during theft detection by the trained data sequence generator SGEN and the first generative machine learning module LLM1.
[0113] The trained data sequence generator SGEN generates test data sequences CS and feeds them into both the generative machine learning module LLM under test and the first generative machine learning module LLM1. Consequently, the generative machine learning module LLM generates a respective test data sequence CS from each CS.
[0114] 16
[0115] Response data set A and a respective response data set CA derived by the first generative machine learning module LLM1.
[0116] The respective response data set A is compared by the detector DET with the respective response data set CA, whereby a respective deviation D between the respective response data sets A and CA is determined.
[0117] The determination of the respective deviation D and its verification by the evaluation module EV can be carried out as described in connection with Figure 2.
[0118] The generation of the test data sequences CS and the testing of the deviations D can continue until a statistical estimate of the probability of error falls below a predefined threshold. A detection signal DS can then be generated and output if the deviations D lie within the range quantified by the stored variance VAR in a statistically significant manner and / or occur significantly more frequently within this range than would be expected based on the stored maximum overlap OLM or the overlaps OL. Otherwise, a complementary signal is generated and output.
Claims
202501217 17 Patent claims 1. Computer-implemented method for configuring a detector (DET) for detecting a first generative machine learning module (LLM1), wherein a) a learning-based data sequence generator (SGEN) generates a large number of training data sequences (TS), which are each fed into the first generative machine learning module (LLM1) and into a different second generative machine learning module (LLM2), from which the first generative machine learning module generates first response data sets (A1) and the second generative machine learning module generates second response data sets (A2), b) a variance (VAR) of a first statistical distribution (SD1) of the first response data sets (A1) and an overlap (OL) of a second statistical distribution (SD2) of the second response data sets (A2) with the first statistical distribution (SD1) are determined, c) the data sequence generator (SGEN) is trained to reduce the variation (VAR) and the overlap (OL), d) a test data sequence (CS) generated by the trained data sequence generator (SGEN) is assigned to a response data set (CA) generated therefrom by the first generative machine learning module (LLM1), and e) the detector (DET) is set up to feed the test data sequence (CS) into a generative machine learning module (LLM) to be tested, to compare a resulting response data set (A) with the response data set (CA) assigned to the test data sequence (CS) and to output a detection signal (DS) depending on the result.
2. The method according to claim 1, characterized in that, that a weighted sum (SC) of the variance (VAR) and the overlap (OL) is formed, and that the training of the data sequence generator (SGEN) is aimed at minimizing the weighted sum (SC).
3. Method according to one of the preceding claims, characterized in that that the determined variance (VAR) is compared with a predetermined first threshold and the determined overlap (OL) with a predetermined second threshold, and that the training will end as soon as the first threshold and the second threshold are both undershot. 202501217 18 4. Method according to one of the preceding claims, characterized in that that the trained data sequence generator (SGEN) generates several test data sequences (CS), that each generated test data sequence (CS) is assigned a response data set (CA) generated from it by the first generative machine learning module (LLM1), and that the detector (DET) is set up to feed the test data sequences (CS) into a generative machine learning module (LLM) to be tested, to compare a resulting response data set (A) with the response data set (CA) assigned to the respective test data sequence (CS) and to output the detection signal (DS) depending on this.
5. Method according to one of the preceding claims, characterized in that that individual training data sequences (TS) are repeatedly fed into the first and second generative machine learning modules, and that the first statistical distribution (SD1) is formed over the first response data sets (A1) of the repeatedly fed training data sets (TS) and the second statistical distribution (SD2) is formed over the second response data sets (A2) of the repeatedly fed training data sets (TS).
6. Method according to one of the preceding claims, characterized in that that the first response data sets (A1) are each encoded by a first vector and the second response data sets (A2) are each encoded by a second vector, that the variance (VAR) is derived from a statistical variance of the first vectors, and that the overlap (OL) is derived from the distances between the first vectors and the second vectors.
7. Method according to one of the preceding claims, characterized in that the overlap (OL) is derived from a Szymkiewicz-Simpson coefficient and / or a Jaccard coefficient of the first and second statistical distribution (SD1, SD2).
8. Method according to one of the preceding claims, characterized in that the data sequence generator (SGEN) is trained using a reinforcement learning method.
9. Method according to one of the preceding claims, characterized in that, 202501217 19 that several second generative machine learning modules (LLM2) are planned, - into which the multitude of training data sets (TS) are fed, and - for which an overlap (OL) of a respective second statistical distribution (SD2) of the second response data sets (A2) with the first statistical distribution (SD1) is determined, and that the data sequence generator (SGEN) is trained to reduce the variation (VAR) and the overlaps (OL).
10. Method according to claim 9, characterized in that, that a maximum overlap (OLM) is selected from the overlaps (OL), and that the data sequence generator (SGEN) is trained to reduce the variation (VAR) as well as the maximum overlap (OLM).
11. Method according to claim 9 or 10, characterized in that, that the overlaps (OL) are added to a total overlap, and that the data sequence generator (SGEN) is trained to reduce the variation (VAR) and the total overlap.
12. Computer-implemented method for theft detection of a generative machine learning module, wherein - a detector (DET) is configured by a method according to one of the preceding claims, - the test data sequence (CS) is fed into a generative machine learning module (LLM) to be tested via the configured detector (DET) and a resulting response data set (A) is read in, - a discrepancy (D) is determined between the read response data set (A) and the response data set (CA) assigned to the test data sequence (CS), and - depending on the determined deviation (D), a detection signal (DS) is output.
13. Theft reporting system (DMS) for a generative machine learning module, configured to execute all process steps of a method according to claim 12.
14. Computer program product comprising instructions which, when executed by a computer, cause the computer to execute a method according to any one of claims 1 to 12. 202501217 20 15. Computer-readable storage medium containing a computer program product according to claim