Anti-money laundering model training method based on federated learning

Abstract

The present application belongs to the anti-money laundering model training method, specifically the anti-money laundering model training method based on federated learning, each participating institution constructs account transfer flow into a continuous time dynamic graph, under the condition that the cloud is consistent with the model structure of each institution, a model for outputting transaction level risk score is constructed; a structure template detection output branch is added on the node embedding and edge embedding output by the shared encoder, the cloud maintains a structure template library, and a structure template evaluation set is constructed based on the template library, each institution optimizes the main task with private transaction data and optimizes the auxiliary task with the structure template evaluation set, and obtains the local model update; and the relative conventional evaluation index improvement and the structure template evaluation index improvement of the global model are calculated, and the cloud determines the qualification according to the two types of improvements, so as to improve the new money laundering structure recognition ability without losing privacy.

Description

Technical Field

[0001] This invention pertains to anti-money laundering model training methods, specifically referring to an anti-money laundering model training method based on federated learning. Background Technology

[0002] Anti-money laundering models are typically trained using federated learning. In federated training with secure aggregation, the cloud cannot obtain the update details of individual institutions. Existing anti-poisoning methods often use clustering followed by outlier removal to filter out anomalous contributions. However, in the early stages of real-world new money laundering, a few institutions may contribute anomalously, which can easily be misfiltered and reduce the ability to identify new types of money laundering.

[0003] During secure aggregation, each node forms an opposite mask with another node. This mask is used to encrypt the node's own parameters. After aggregation in the cloud, the masks cancel each other out. Although the cloud cannot see the actual parameters of each node, it can still obtain the correct aggregated data. However, the parameters are encrypted by the mask, making it impossible to locate malicious participants and exclude the training data they send.

[0004] In summary, existing federated learning model training methods struggle to identify malicious actors, and even when outlier parameters are removed, they fail to retain the ability to identify new money laundering methods. Summary of the Invention

[0005] The purpose of this invention is to provide a method for training anti-money laundering models based on federated learning, so as to solve the above-mentioned technical problems.

[0006] The method for training an anti-money laundering model based on federated learning according to the present invention includes the following steps.

[0007] S1, Each participating institution will construct a continuous time dynamic graph of its own account transfer records;

[0008] S2, Under the consistent model structure of each participating institution and the cloud server, a transaction-level alarm model is established. The transaction-level alarm model uses a time-series graph neural network to encode the event sequence of the continuous time dynamic graph and output a transaction-level risk score.

[0009] S3, add a structural template detection output branch on top of the node embedding and edge embedding output by the shared encoder of the temporal graph neural network, and output the structural template matching score or template category probability distribution;

[0010] S4, cloud server or alliance maintains the structural template library and constructs structural template evaluation set;

[0011] S5, each participating institution performs local training, optimizes the main task on private transaction data and optimizes the auxiliary task on the structure template evaluation set, and obtains local model updates;

[0012] S6, Each participating institution shall calculate and report the improvement of the regular assessment indicators and the improvement of the structural template assessment indicators, and shall not upload the plain text of the parameter update for each institution.

[0013] S7, The cloud server determines the eligibility of participants and generates a list of qualified participants based on the improvement of the conventional evaluation indicators and the improvement of the structural template evaluation indicators;

[0014] S8, during the r-th round of aggregation, the cloud server determines the participating set for this round based on the qualification list formed in the (r-1)-th round, executes the secure aggregation protocol on the participating set to obtain the aggregation update and update the global model.

[0015] Furthermore, the continuous-time dynamic graph includes a node table with account ID as the key and an edge event table with event ID as the key.

[0016] Furthermore, the continuous-time dynamic graph constructs an event-induced subgraph index through a sliding time window Tw, and extracts a k-hop neighborhood subgraph centered on the target transaction or an event sequence fragment truncated by the sliding time window Tw for the output branch of the structure template detection.

[0017] Furthermore, the temporal graph neural network includes a time encoder, a graph message calculation unit, a memory update unit, and a readout prediction unit;

[0018] The memory update unit maintains a memory vector of length d for each account and updates the memory vector when a transfer event occurs.

[0019] Furthermore, each structural template in the structural template library consists of a pattern type and a parameter set. The pattern type includes at least splitting, converging, looping, and layered posting. The parameter set includes at least the fan-out or fan-in degree range, the number of layers range, the amount ratio range, the time interval range, and the noise tolerance.

[0020] Furthermore, the structural template evaluation set consists of public or synthetic samples, and the synthetic samples are generated in at least one of the following ways:

[0021] The structural templates in the aforementioned structural template library are subjected to parameterized perturbations based on the amount ratio, time interval, number of layers, and degree.

[0022] Alternatively, a synthetic event sequence and its template labels that satisfy the constraints of the structure template library can be generated through a rule engine or simulator.

[0023] Furthermore, the local training employs weighted loss for multi-task joint optimization:

[0024]

[0025] in This represents the task weight.

[0026] Furthermore, the task weight according to and The moving average ratio is adaptively adjusted and limited to Within the range.

[0027] Furthermore, the participation eligibility determination is based on a structural template threshold. Conventional degradation threshold And allow for a short-term decrease in the lower limit A joint determination will be made.

[0028] Furthermore, the secure aggregation protocol includes: each participating institution uploads its local model update vector after adding a mask, wherein the mask is the sum of pairwise masks generated by pairwise random seeds of other participating parties.

[0029] The beneficial effects achieved by the present invention using the above solution are as follows:

[0030] The training method of this invention, under the premise of secure aggregation without disclosing the local transaction data of each institution and the update details of the single institution model, can distinguish between performance damage caused by malicious poisoning and a few abnormal contributions in the early stage of new money laundering through dual-domain evaluation and second-round qualification control elimination mechanism, thereby achieving effective isolation of destructive participating institutions and protecting new money laundering signals from being falsely filtered. Attached Figure Description

[0031] Figure 1 This is a flowchart of the overall anti-money laundering model based on federated learning of the present invention.

[0032] Figure 2 This is a schematic diagram of the structure of the local neural network of the present invention. Detailed Implementation

[0033] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0034] In this embodiment, the alliance includes multiple participating institutions and a cloud server. Each participating institution trains a transaction-level alarm model based on federated learning and secure aggregation without exposing its private transaction details to the cloud server. It also removes destructive updates through dual-domain evaluation and second-round eligibility control, while protecting its ability to learn new money laundering structure signals.

[0035] The anti-money laundering model training method based on federated learning proposed in this invention includes the following steps:

[0036] S1, Each participating institution constructs a continuous time dynamic graph of its account transfer records, where nodes are accounts and edges are timestamped transfer events, and features are configured for nodes and edges;

[0037] S2, Under the same model structure of each participating institution and the cloud server, a transaction-level alarm model is established for outputting transaction-level risk scores. The transaction-level alarm model uses a time-series graph neural network to encode the event sequence of the continuous time dynamic graph.

[0038] S3, on top of the node embedding and edge embedding output by the shared encoder of the temporal graph neural network, a structural template detection output branch is added to output a structural template matching score;

[0039] S4, maintain the structure template library in the cloud, and construct a structure template training set and a structure template evaluation set that do not contain single-mechanism private real flow based on the structure template library;

[0040] S5, each participating institution performs local training locally, optimizes the main task on private transaction data and optimizes the auxiliary task on the structure template training set to obtain local model updates;

[0041] S6, each participating institution calculates the improvement of the conventional evaluation index ΔI and the improvement of the structural template evaluation index ΔI on the conventional validation set and the structural template evaluation set respectively, relative to the global model issued in this round, and reports the improvement to the cloud server without uploading the single institution parameter update / gradient plaintext.

[0042] S7, the cloud server determines the eligibility of participants based on the improvement of the conventional evaluation indicators and the improvement of the structural template evaluation indicators, and writes the determination result into the eligibility list for the next round.

[0043] S8, during the r-th round of aggregation, the cloud server determines the participating set for this round based on the qualification list formed in the (r-1)-th round, and executes the secure aggregation protocol only on each participating institution within the participating set to obtain the aggregation update and update the global model.

[0044] Specifically, each participating institution extracts event streams from its own account transfer records and constructs them into a continuous-time dynamic graph data structure.

[0045] The event stream is arranged in ascending order of timestamps, and each transfer event record includes at least: source account ID, target account ID, transfer amount, timestamp, direction identifier, transaction channel / device identifier, account attribute fields (such as account type, account opening duration, and region tag), and adjacent event time difference fields.

[0046] Each participating institution maintains a node table and an edge event table locally: The node table stores the static feature vector of an account using the account ID as the key; the edge event table stores the edge feature vector and its occurrence timestamp using the event ID as the key, and establishes an "account ID → event sequence index" to support querying historical events within a nearby time window by account.

[0047] To facilitate subsequent structural template matching, the participating institutions can also construct an event-induced subgraph index locally according to the sliding time window 𝑇𝑤, which is used to quickly extract a k-hop neighborhood subgraph centered on a certain transaction; where 𝑇𝑤 is the preset time window length and k is the preset neighborhood hop number.

[0048] Each participating organization and the cloud server pre-agreed upon the same model structure and parameter dimensions.

[0049] The transaction-level alarm model uses a time-series graph neural network to encode a continuous time dynamic graph event sequence and output a transaction-level risk score.

[0050] The temporal graph neural network includes: a time encoder, a graph message computation unit, a memory update unit, and a readout prediction unit; wherein, the time encoder is used to map the event occurrence time and the historical time difference into time features; the graph message computation unit is used to aggregate the historical neighborhood information of the source account and the target account when the event is triggered; the memory update unit is used to maintain a length of [missing information] for each account. The memory vector is updated upon event arrival; the readout prediction unit outputs the transaction-level risk score based on the embedding vector corresponding to the transaction. .

[0051] Specifically, a TGN framework containing the memory update unit or a similar continuous-time dynamic graph neural network can be used to handle the characteristics of dynamic graphs evolving over time.

[0052] A structural template detection output branch is added above the node embedding and edge embedding output by the shared encoder. The structural template detection output branch is connected to the shared encoder and shares its input features. It is used to perform structural template matching on the local subgraph / event sequence related to the target transaction. The structural template detection output branch can be trained on the structural template training set.

[0053] The input to the structure template detection output branch is: a k-hop neighborhood local subgraph representation extracted within the sliding time window 𝑇𝑤 centered on the target transaction event, and the node embedding sequence and edge embedding sequence corresponding to the local subgraph;

[0054] The output of the structural template detection output branch is a structural template matching score vector 𝐩=(p1,…,pM) or a template category probability distribution, where M is the number of templates in the structural template library, used to represent the degree of matching between the local subgraph / event sequence and one or more templates in the structural template library.

[0055] This results in multi-task output:

[0056] The main task output is the transaction-level risk score;

[0057] The auxiliary task output is the structure template matching score.

[0058] The cloud server or alliance maintains the structure template library, which contains at least parametric descriptions of several subgraph patterns.

[0059] Each structural template can be composed of "mode type + parameter set": the mode type includes splitting, aggregation, looping, layer-by-layer posting, etc.; the parameter set includes at least the fan-out / fan-in degree range, the number of layers range, the amount ratio range, the time interval range, and the tolerance for missing edges / noise.

[0060] The training and evaluation sets of the structural templates consist of public or synthetic samples, and neither contains any single-institution private real-world pipelines.

[0061] Firstly, a new variant sample can be generated by abstracting a known typical money laundering structure into a template and then perturbing the amount ratio, time interval, number of layers and degree.

[0062] Secondly, a synthetic event sequence and corresponding labels that satisfy template constraints can be randomly generated by a rule engine or simulator.

[0063] Preferably, the above-mentioned public or synthetic samples are divided into a structural template training set and a structural template evaluation set according to a preset ratio. The structural template training set is used for training the auxiliary task, and the structural template evaluation set is used to calculate the improvement amount ΔR of the structural template evaluation index and to participate in the qualification determination in the cloud, so as to evaluate the model's sensitivity and recognition ability to new structural templates.

[0064] At the start of each training round, the cloud server distributes global model parameters to each participating institution. Each participating institution optimizes the main task on its local private transaction data, and simultaneously optimizes the auxiliary task on the structure template training set or a subset thereof, using weighted loss for joint optimization.

[0065]

[0066] in, It can be the transaction-level binary cross-entropy or the loss with class imbalance weights; It can be template category cross-entropy or multi-label matching loss; This represents the task weight.

[0067] To improve stability, An adaptive update method can be used. Preferably, at the end of each round, the update is based on the ratio of the moving averages of the two types of losses. This keeps the gradient magnitudes of the two tasks within a preset ratio range;

[0068] In another embodiment, 𝜆 can be limited to [ , The system is adjusted slowly using an exponential moving average method to avoid a significant short-term degradation of the main task due to excessive auxiliary tasks.

[0069] After completing local training, each participating institution does not upload plaintext parameter updates. Instead, they calculate two types of improvements relative to the global model distributed in this round and report them to the cloud server.

[0070] 1. Improvement in regular evaluation metrics Δ𝑄𝑖: The difference in regular evaluation metrics on a pre-agreed regular validation set. The difference is the metric value of the "locally updated model" minus the metric value of the "global model issued in this round".

[0071] Conventional evaluation metrics should preferably be those that are as large as possible (including but not limited to AUC and F1). When using a loss function, the amount of loss reduction should be converted into the amount of positive improvement.

[0072] The standard validation set is a public / consortium shared dataset, and is decoupled from single-organization private data;

[0073] 2. Improvement of structural template evaluation index Δ𝑅𝑖: The difference in template identification index on the structural template evaluation set, wherein the difference is the template identification index value of the "locally updated model" minus the template identification index value of the "global model issued in this round";

[0074] Template recognition metrics can be the positive improvement calculated from template classification accuracy, recall, macro average F1, or the decrease in matching loss.

[0075] The cloud server uses (Δ𝑄𝑖, Δ𝑅𝑖) to conduct the next round of qualification assessment for participating institutions, and writes the assessment results into the qualification list for the next round.

[0076] The eligibility criteria only apply to the next round of participants to ensure that the current round's secure aggregation protocol is computable;

[0077] Optionally, the cloud server performs range constraints and historical consistency checks on the reported Δt and Δt, and those exceeding the preset reasonable range are included in the observation set or isolation list.

[0078] Thresholds include: structural template thresholds Conventional degradation threshold And allow for a short-term decrease in the lower limit .

[0079] The threshold update method can adopt a sliding window adaptive approach: the cloud server maintains the most recent K rounds of "retained institutions". and Statistics, with thresholds updated based on quantiles;

[0080] For example: Let Take the nearest wheel The quantiles, let Pick The quantiles, and let Pick Where 𝐾 is a positive integer, 0 <p<1,0<q<1,β∈(0,1)。

[0081] The following is an example of the judgment rule:

[0082] New type of effective contribution determination: when and Not less than When this occurs, it is marked as a new type of valid contribution, retains eligibility for the next round, and its sampling probability or weight cap can be increased;

[0083] Destructive update determination: When and If an update is flagged as disruptive, it will be added to the quarantine list and disqualified from the next round.

[0084] Observation criteria: All other cases are added to the observation set. The cloud server can introduce continuous... Round triggering mechanism, i.e., continuous Isolation is only implemented when the destructive update criteria are met.

[0085] In the During round aggregation, the cloud server uses the first... The list of qualified participants for this round is determined by the list of participants formed in this round. Only for Internal organizations execute secure aggregation protocols.

[0086] Each participating institution updates its model vector based on the model obtained through local training. Upload after masking: The mask can be composed of the sum of pairs of masks generated by pairs of random seeds from other participants, and the mask can be split into several mask fragments to adapt to the disconnection recovery mechanism.

[0087] The cloud server only recovered through security aggregation. And unable to obtain any single institution Plain text, then according to learning rate Update global model parameters.

[0088] Since the removal occurs in the next round of participation eligibility, the protocol will not become unsolvable due to temporary removal in this round of security aggregation; the next round of aggregation excludes destructive entities at the set level, and the aggregation update obtained by the cloud server is the true aggregation result of the qualified set, thereby achieving continuous removal of malicious entities and continuous protection of new structural signals without sacrificing privacy.

[0089] The above description is merely a specific embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications, equivalent substitutions, or improvements made to the above embodiments without departing from the technical concept of the present invention should fall within the protection scope of the present invention. The protection scope of the present invention is determined by the scope defined in the claims, and the specification and drawings can be used to interpret the claims.

Claims

1. A method for training an anti-money laundering model based on federated learning, characterized in that, Under safe aggregation conditions, cross-round participation eligibility control is implemented based on the improvement of structural template evaluation indicators and the improvement of conventional evaluation indicators, including the following steps: S1, Each participating institution will construct a continuous time dynamic graph of its own account transfer records; S2, Under the consistent model structure of each participating institution and the cloud server, a transaction-level alarm model is established. The transaction-level alarm model uses a time-series graph neural network to encode the event sequence of the continuous time dynamic graph and output a transaction-level risk score. S3, add a structural template detection output branch on top of the node embedding and edge embedding output by the shared encoder of the temporal graph neural network, and output the structural template matching score or template category probability distribution; S4, cloud server or alliance maintains the structural template library and constructs structural template evaluation set; S5, each participating institution performs local training, optimizes the main task on private transaction data and optimizes the auxiliary task on the structure template evaluation set, and obtains local model updates; S6, Each participating institution calculates the improvement of the conventional evaluation indicators on the conventional validation set. And calculate the improvement of the structural template evaluation index on the structural template evaluation set. and report to the cloud server. and Furthermore, it does not upload plaintext updates of single-organization parameters; The regular validation set is a public dataset; S7, the cloud server determines the eligibility of participants based on the improvement of the conventional evaluation indicators and the improvement of the structural template evaluation indicators, and generates a list of eligible participants for the next round of aggregation. S8, during the r-th round of aggregation, the cloud server determines the participating set for this round based on the qualification list formed in the (r-1)-th round, and executes the secure aggregation protocol only on the participating institutions within the participating set to obtain aggregation updates and update the global model.

2. The method for training an anti-money laundering model based on federated learning according to claim 1, characterized in that, The continuous-time dynamic graph includes a node table with account ID as the key and an edge event table with event ID as the key.

3. The method for training an anti-money laundering model based on federated learning according to claim 1, characterized in that, The continuous-time dynamic graph constructs an event-induced subgraph index through a sliding time window Tw, and extracts a k-hop neighborhood subgraph centered on the target transaction or an event sequence fragment truncated by the sliding time window Tw for the output branch of the structure template detection.

4. The method for training an anti-money laundering model based on federated learning according to claim 3, characterized in that, The temporal graph neural network includes a time encoder, a graph message computation unit, a memory update unit, and a readout prediction unit; The memory update unit maintains a memory vector of length d for each account and updates the memory vector when a transfer event occurs.

5. The method for training an anti-money laundering model based on federated learning according to claim 1, characterized in that, Each structural template in the structural template library consists of a pattern type and a parameter set. The pattern type includes at least splitting, converging, looping, and layered posting. The parameter set includes at least the fan-out or fan-in degree range, the number of layers range, the amount ratio range, the time interval range, and the noise tolerance.

6. The method for training an anti-money laundering model based on federated learning according to claim 1, characterized in that, The structural template evaluation set consists of public or synthetic samples, and the synthetic samples are generated in at least one of the following ways: The structural templates in the aforementioned structural template library are subjected to parameterized perturbations based on the amount ratio, time interval, number of layers, and degree. Alternatively, a synthetic event sequence and its template labels that satisfy the constraints of the structure template library can be generated through a rule engine or simulator.

7. The method for training an anti-money laundering model based on federated learning according to claim 1, characterized in that, The local training employs weighted loss for multi-task joint optimization. in This represents the task weight.

8. The method for training an anti-money laundering model based on federated learning according to claim 7, characterized in that, The task weight according to and The moving average ratio is adaptively adjusted and limited to Within the range.

9. The method for training an anti-money laundering model based on federated learning according to claim 1, characterized in that, The participation eligibility determination is based on a structural template threshold. Conventional degradation threshold And allow for a short-term decrease in the lower limit Joint determination is performed, including: when and Not less than At that time, the corresponding participating institutions will be added to the list of qualified institutions for the next round; when and If so, the corresponding participating institution will be removed from the list of qualified institutions in the next round.

10. The method for training an anti-money laundering model based on federated learning according to claim 1, characterized in that, The secure aggregation protocol includes: each participating institution uploads its local model update vector after adding a mask, wherein the mask is the sum of pairwise masks generated by pairwise random seeds of other participating parties.

Images