AMLNet: Dual-Use AML & Forecasting

• AMLNet is an overloaded term representing two distinct systems: one for regulation-aligned synthetic AML transaction generation and detection, and another for neural time series forecasting.
• The anti-money laundering variant simulates realistic transactions using agent-based generation and an ensemble detection pipeline that achieves high precision and recall.
• The forecasting variant employs auto-regressive and non-auto-regressive decoders with online knowledge distillation, delivering competitive accuracy with reduced latency.

AMLNet is a name used for two distinct research systems in the arXiv literature. In anti-money laundering research, AMLNet denotes a knowledge-based multi-agent framework that generates regulation-aligned synthetic transactions and applies an ensemble detection pipeline to identify laundering behavior. In time-series forecasting, AMLNet denotes an adversarial mutual learning neural network for non-auto-regressive multi-horizon forecasting that couples auto-regressive and non-auto-regressive decoders through online knowledge distillation (Huda et al., 15 Sep 2025, Lin, 2023).

1. Disambiguation of the term

The shared label “AMLNet” does not designate a single canonical architecture. One usage belongs to financial crime research and centers on synthetic anti-money laundering data generation and detection. The other belongs to sequence modeling and addresses multi-horizon time series forecasting.

Domain	Full title	Core function
Anti-money laundering	“AMLNet: A Knowledge-Based Multi-Agent Framework to Generate and Detect Realistic Money Laundering Transactions”	Synthetic transaction generation and laundering detection
Time-series forecasting	“AMLNet: Adversarial Mutual Learning Neural Network for Non-AutoRegressive Multi-Horizon Time Series Forecasting”	NAR multi-horizon forecasting with online KD

A common misconception is to treat references to “AMLNet” as unambiguous. The literature shows that the name is overloaded across unrelated problem settings, model classes, and evaluation protocols. This suggests that citations and discussions benefit from explicit domain qualification, such as “AMLNet for AML” or “AMLNet for forecasting.”

2. Multi-agent AMLNet for synthetic anti-money laundering research

In the anti-money laundering setting, AMLNet comprises two tightly coupled units: a Transaction Generation Unit and a Money Laundering Detection Unit (Huda et al., 15 Sep 2025). The Transaction Generation Unit contains Customer Simulation Agents, Transaction Generation Agents, and AML Pattern Injection Agents. Customer Simulation Agents instantiate realistic customer profiles and spending patterns. Transaction Generation Agents invoke each Customer Simulation Agent’s simulate_day() to produce normal transactions. AML Pattern Injection Agents overlay suspicious behaviors, including structuring, layering, and integration, on selected Customer Simulation Agents.

The Money Laundering Detection Unit contains Detection Agents and Real-time Monitoring Agents. Detection Agents extract amount, temporal, and network features and apply an ensemble of Isolation Forest and Random Forest models. Real-time Monitoring Agents continuously update risk assessments and generate alerts. A human-in-the-loop feedback loop feeds detection metrics and expert corrections back into generator rule parameters, with the stated purpose of ensuring iterative refinement toward regulatory alignment and technical fidelity.

The high-level workflow is defined as follows. Customer Simulation Agents are initialized using demographic and risk profiles and connected into a transaction network $G(V,E)$. For each day, each agent runs simulate_day(), and Transaction Generation Agents produce normal transactions. With probability $suspicious\_rate$, an AML Pattern Injection Agent overrides Customer Simulation Agent behavior to inject structuring, layering, or integration. All labeled transactions are then processed by the detection stack through feature extraction, SMOTE plus undersampling, Isolation Forest plus Random Forest, risk scoring, and alerting. Detection performance and expert review are subsequently used to update Customer Simulation Agent and AML Pattern rules.

3. Regulation-aware generation and fidelity criteria

The generator is explicitly knowledge-based and regulation-aware. AUSTRAC rules, including the reporting threshold of $10 000 and typology definitions, are encoded as decision rules in AML Pattern Injection Agents [2509.11595]. Customer Simulation Agents sample spending profiles via a Dirichlet distribution,$$\mathbf{p}\sim\mathrm{Dirichlet}([22.1,20.0,16.9,15.0,13.0,5.9,4.0,3.0,0.1] ),$$ where each parameter maps to one of 11 spending categories, including Housing, Food, Transport, and Shell Company. Transaction timing follows business-hour and monthly salary-cycle heuristics based on Australian ABS and RBA data.

Laundering phases are modeled at multiple sophistication levels. For structuring, low sophistication splits an amount into equal parts each below $9 500; medium sophistication uses a$\mathrm{Normal}(\mu,\sigma)$split clipped below $9 500; high sophistication samples split proportions $\mathbf{s}\sim\mathrm{Dirichlet}(\mathbf{\alpha})$, enforces 1–2 transactions in the $8 500–$9 900 range, and generates the remainder via LogNormal. For layering, low sophistication uses 2–3 layers with fixed delays; medium sophistication uses Dirichlet splits and randomized delays of 1–3 days; high sophistication uses 5–8 layers, irregular delays, account reuse, and parallel paths. For integration, low sophistication uses a single payment of at least $20 000 to a shell; medium sophistication uses payment to merchant or crypto with legitimacy score; high sophistication uses multiple payments over weeks into property, luxury goods, and related targets. Advanced behaviors include adaptive threshold behavior.

Rule-matching functions map each injected transaction to AUSTRAC typologies, and all 1 745 suspicious transactions map to at least one of six AUSTRAC patterns. The alignment metric assigns 1 point if a generated percentage falls within AUSTRAC’s expected range, 0.5 if the pattern is present but outside range, and 0 if absent, producing

$$\mathrm{Alignment} = \frac{\sum_{i=1}^6\mathrm{points}_i}{6}\times100\% = 75\%.$$

Pattern-level results are Placement/Structuring at 18.4% and well-aligned, Layering at 78.5% and over-represented, Integration at 3.1% and under-represented, Unusual Volumes at 84.9% and over-represented, High-Risk Categories at 5.4% and aligned, and Unusual Account Activity at 7.7% and aligned.

Technical fidelity is evaluated across temporal, structural, and behavioral dimensions. Temporal realism $S_t$ is measured by DTW against RBA/ABS reference patterns; structural realism $suspicious\_rate$0 is measured by Graph Edit Distance against a reference financial network; behavioral realism $suspicious\_rate$1 combines category, risk, alert, and fraud-rate metrics:

$suspicious\_rate$2

with $suspicious\_rate$3, $suspicious\_rate$4, $suspicious\_rate$5, and $suspicious\_rate$6, yielding

$suspicious\_rate$7

Using $suspicious\_rate$8, $suspicious\_rate$9, and $10 000 and typology definitions, are encoded as decision rules in AML Pattern Injection Agents [2509.11595]. Customer Simulation Agents sample spending profiles via a Dirichlet distribution,$0, and substituting $10 000 and typology definitions, are encoded as decision rules in AML Pattern Injection Agents [2509.11595]. Customer Simulation Agents sample spending profiles via a Dirichlet distribution,$1, $10 000 and typology definitions, are encoded as decision rules in AML Pattern Injection Agents [2509.11595]. Customer Simulation Agents sample spending profiles via a Dirichlet distribution,$2, and $10 000 and typology definitions, are encoded as decision rules in AML Pattern Injection Agents [2509.11595]. Customer Simulation Agents sample spending profiles via a Dirichlet distribution,$3, the composite technical fidelity score is 0.75.

4. Detection ensemble, dataset release, and stated limitations

The AMLNet detection pipeline operates on batches of transactions, exemplified by 1 000 transactions per batch, and extracts amount, temporal, and network features (Huda et al., 15 Sep 2025). It then applies SMOTE plus RandomUnderSampler, computes anomaly scores with Isolation Forest, produces risk scores in $10 000 and typology definitions, are encoded as decision rules in AML Pattern Injection Agents [2509.11595]. Customer Simulation Agents sample spending profiles via a Dirichlet distribution,$4 with a Random Forest Classifier, flags transactions when $10 000 and typology definitions, are encoded as decision rules in AML Pattern Injection Agents [2509.11595]. Customer Simulation Agents sample spending profiles via a Dirichlet distribution,$5, and incrementally updates models. The feature set contains nine features in total. Amount features are absolute amount, relative size, and structuring indicators. Temporal features are velocity, periodicity, and deviation from expected. Network features are degree, centrality, and clustering coefficient. Fusion combines Isolation Forest anomalies and Random Forest risk probabilities via a threshold $10 000 and typology definitions, are encoded as decision rules in AML Pattern Injection Agents [2509.11595]. Customer Simulation Agents sample spending profiles via a Dirichlet distribution,$6 tuned to balance precision and recall.

The released dataset contains 1 090 173 transactions over 195 days, of which 1 745 are laundering-positive, corresponding to 0.16%. The typologies covered are placement, structuring, layering, and integration at low, medium, and high sophistication, together with advanced behaviors such as adaptive thresholds. On internal AMLNet test partitions, the ensemble attains precision 0.84, recall 0.97, F1 0.90, and ROC-AUC 0.88. On the external SynthAML dataset, the reported values are ROC-AUC 0.80, precision 0.67, recall 0.72, and F1 0.69. The paper interprets this as indicating architectural generalizability across different synthetic generation paradigms.

The dataset is released as Version 1.0 on Zenodo at https://doi.org/10.5281/zenodo.16736515 under CC BY-NC 4.0, and the code is to be published on GitHub. Recommended uses are benchmarking and algorithm development in AML research and validation of new detection models in a regulation-aligned synthetic environment. The stated limitations are that integration patterns are under-represented, rare typologies are simplified, and the data may not capture all jurisdictional nuances or re-identification risks. The paper also states that the dataset is not a replacement for institution-specific real data and that local validation and governance are required.

5. AMLNet as an adversarial mutual learning network for forecasting

In the forecasting literature, AMLNet denotes a neural architecture for non-auto-regressive multi-horizon time series forecasting (Lin, 2023). The model consists of a shared encoder $10 000 and typology definitions, are encoded as decision rules in AML Pattern Injection Agents [2509.11595]. Customer Simulation Agents sample spending profiles via a Dirichlet distribution,$7 of depth $10 000 and typology definitions, are encoded as decision rules in AML Pattern Injection Agents [2509.11595]. Customer Simulation Agents sample spending profiles via a Dirichlet distribution,$8, two “deep” peer decoders of equal depth $10 000 and typology definitions, are encoded as decision rules in AML Pattern Injection Agents [2509.11595]. Customer Simulation Agents sample spending profiles via a Dirichlet distribution,$9, and one “shallow” non-auto-regressive student decoder $$\mathbf{p}\sim\mathrm{Dirichlet}([22.1,20.0,16.9,15.0,13.0,5.9,4.0,3.0,0.1] ),$$0 of depth $$\mathbf{p}\sim\mathrm{Dirichlet}([22.1,20.0,16.9,15.0,13.0,5.9,4.0,3.0,0.1] ),$$1. Peer1 ($$\mathbf{p}\sim\mathrm{Dirichlet}([22.1,20.0,16.9,15.0,13.0,5.9,4.0,3.0,0.1] ),$$2) is auto-regressive, Peer2 ($$\mathbf{p}\sim\mathrm{Dirichlet}([22.1,20.0,16.9,15.0,13.0,5.9,4.0,3.0,0.1] ),$$3) is non-auto-regressive, and the student is also non-auto-regressive. All decoders consume the encoder’s last hidden states $$\mathbf{p}\sim\mathrm{Dirichlet}([22.1,20.0,16.9,15.0,13.0,5.9,4.0,3.0,0.1] ),$$4 and a fixed start-token window of covariates $$\mathbf{p}\sim\mathrm{Dirichlet}([22.1,20.0,16.9,15.0,13.0,5.9,4.0,3.0,0.1] ),$$5, while $$\mathbf{p}\sim\mathrm{Dirichlet}([22.1,20.0,16.9,15.0,13.0,5.9,4.0,3.0,0.1] ),$$6 additionally uses previous ground truths.

Forecasts are modeled distributionally. Each output $$\mathbf{p}\sim\mathrm{Dirichlet}([22.1,20.0,16.9,15.0,13.0,5.9,4.0,3.0,0.1] ),$$7 is a Gaussian with timestep-wise mean $$\mathbf{p}\sim\mathrm{Dirichlet}([22.1,20.0,16.9,15.0,13.0,5.9,4.0,3.0,0.1] ),$$8 and variance $$\mathbf{p}\sim\mathrm{Dirichlet}([22.1,20.0,16.9,15.0,13.0,5.9,4.0,3.0,0.1] ),$$9. The base loss for each decoder is the negative log-likelihood:

$9 500; medium sophistication uses a$0

The model’s distinctive feature is online knowledge distillation trained from scratch. Outcome-driven knowledge distillation dynamically re-weights each sample’s KL distillation loss by how well the teacher fitted that sample. The outcome weight is defined as

$9 500; medium sophistication uses a$1

Peer losses and student losses use this weighting to distill from the deep AR and NAR branches. Hint-driven knowledge distillation aligns hidden-state distributions adversarially, motivated by the claim that AR decoder hidden states are smoother and encode output-space continuity.

6. Distillation mechanics, training protocol, and empirical behavior

Hint-driven distillation uses a discriminator for each of the $9 500; medium sophistication uses a$2 hidden layers in $9 500; medium sophistication uses a$3 and $9 500; medium sophistication uses a$4 (Lin, 2023). Each discriminator takes a hidden-state tensor $9 500; medium sophistication uses a$5 and outputs a scalar in $9 500; medium sophistication uses a$6. The discriminator architecture is specified as Conv1d with output 16, kernel 3, stride 2, followed by BatchNorm and LeakyReLU, then Conv1d with output 1, kernel 3, stride 1, followed by Linear and Sigmoid. Decoder losses use adversarial objectives against peer discriminators, and discriminator losses distinguish peer hidden states from the other peer and the student.

The training algorithm proceeds per minibatch by encoding inputs, forwarding $9 500; medium sophistication uses a$7 and $9 500; medium sophistication uses a$8, updating the encoder and peer parameters with negative log-likelihood plus outcome-driven and hint-driven losses, forwarding the student, updating student parameters, updating discriminator parameters, and repeating until convergence. At test time, only $9 500; medium sophistication uses a$9 are used. Implementation details are given as Tesla V100, PyTorch 1.6, Adam, 200 epochs, and Bayesian hyperparameter search.

Evaluation uses the Sanyo, Hanergy, Solar, and Electricity datasets. The splits are defined as last year or week for test, preceding year or week for validation, and the remainder for training. Baselines are Persistence, SARIMAX, DeepAR, LogSparse-Trans (“LogTrans”), N-BEATS, and Informer. Metrics are the $split clipped below $0 and $split clipped below $1 quantile losses. Selected reported results are as follows.

Model	Sanyo 0.5/0.9	Hanergy 0.5/0.9
Informer	0.046 / 0.022	0.084 / 0.046
AMLNet-P1	0.044 / 0.021	0.084 / 0.043
AMLNet-P2	0.040 / 0.019	0.078 / 0.040
AMLNet-S	0.042 / 0.020	0.077 / 0.038

Inference-time comparisons on Sanyo and Hanergy report approximately $split clipped below $2 ms and $split clipped below $3 ms per sample for LogTrans, 18.1 $split clipped below $4 0.5 and 18.7 $split clipped below $5 1.1 for Informer, 21.7 $split clipped below $6 0.2 and 21.5 $split clipped below $7 0.1 for AMLNet-P2, and 11.4 $split clipped below $8 0.5 and 11.6 $split clipped below $9 0.5 for AMLNet-S. The paper states that AMLNet-S matches or beats Informer in accuracy while running approximately 1.6× faster.

Ablation results attribute limited benefit to classic online KD, marginal benefit to adversarial distillation on the last layer, a sizable gain to outcome-driven KD, and the best results to the full combination of outcome-driven KD and hidden-state GAN distillation. The horizon study reports that the $\mathbf{s}\sim\mathrm{Dirichlet}(\mathbf{\alpha})$0 decoder degrades much more slowly than vanilla AR (DeepAR), while the NAR branches remain stable. Hidden-state cosine distances drop by approximately 23–28% versus Informer, and DTW path-distances drop approximately 17–18%, which the paper presents as evidence of smoother feature dynamics.

7. Relationship between the two AMLNet systems

The two AMLNet systems share a title string but not a methodological substrate. The anti-money laundering AMLNet is a knowledge-based multi-agent framework with explicit rules, transaction-network simulation, regulatory alignment scoring, and a hybrid anomaly-risk detection stack. The forecasting AMLNet is a neural sequence model centered on AR/NAR peer decoders, online KL-based distillation, adversarial hint transfer, and low-latency inference.

Their juxtaposition is nevertheless informative. Both systems combine heterogeneous components rather than relying on a single homogeneous model family: generator plus detector in one case, and AR plus NAR peers plus a shallow student in the other. This suggests a broader pattern in contemporary machine learning nomenclature, where identical model names may emerge independently in different subfields while denoting structurally distinct research programs.