Federated Learning в iGaming

1) Why FL is in iGaming

Federated training (FL) allows multiple participants (brands, regions, providers, PSPs) to train a common model without sharing raw data. This is critical where there is PII/finance, cross-border restrictions and a wide partner perimeter.

• Improving the quality of models due to the "general intelligence" of the holding/partners.
• Reduce legal risks and anonymization/exchange costs.
• Quickly reach new regions without migrating data histories.

Typical tasks: Responsible Gaming (RG) scoring, anti-fraud/chargeback, AML patterns, KYC verification (thin-file), personalization/CRM, bot traffic detection/boosting.

2) FL architectures

Cross-silo (between organizations/brands/regions): a little "smart" participants, stable communication, long sessions. Suitable for holdings/PSP/providers.
Cross-device (many players' devices): millions of "thin" clients, intermittent communication. Less common for iGaming (chats/application clients), but possible for on-device signals.

• Centralized coordinator (server-aggregator) - basic option.
• Hierarchical (regional aggregators → central) - reduces traffic/latency.
• Peer-to-peer/secure aggregation mesh is more difficult, but above the "zero-trust" property.

3) Protection of privacy and security

Secure Aggregation: The server sees only the sum/average of the gradients, not the updates of a particular participant.
Differential privacy (DP): noise on the client side and/or in aggregation; keep records of the ε budget.
Confidential computing (TEE): aggregation and/or inference in isolated enclaves.
MPC/PSI: secure intersections/co-inference calculations with PSP/providers.
Access and logging policies: prohibition on serialization of raw features/gradients; aggregates and metadata only.

4) FL Technical Challenges and How to Solve Them

Non-IID and imbalance: domain data is different (countries, payment methods, lenders).
→ Use personalization over global model (fine-tuning/adapter layers), stratified batches, weighted aggregation (by quality/size).

Heterogeneous hardware/network: participants with different power and availability.
→ Partial participation, asynchronous aggregation, adaptive update sizes.

Compression and traffic: heavy weights/gradients.
→ Quantization, sparsification, sketch coding; less often - transmission of "deltas" instead of full weights.

Poisoning/backdoors: A malicious participant spoils the model.
→ Robust aggregators (median/Krum/trimmed mean), anomaly detectors in updates, honeypot tasks and test sets, reputational weights.

Drift and regression: behavior changes/regulators.
→ Continuous learning, periodic re-init, champion-challenger, ML-observability by segment.

5) Patterns for key cases

5. 1 RG scoring (responsible game)

Goal: Equal Opportunity (do not miss risic players in any country/segment).
Approach: cross-silo FL between brands/regional teams; Secure Agg + DP; local calibration of thresholds.
Overrides: Self-exclusion/limit flags dominate the model.

5. 2 Anti-fraud/payments/chargeback

Target: Equalized Odds (FPR control), resistance to new fraud.
Approach: joint FL between holding operators and PSP; TEE aggregator; MPC for co-inference on payment.
Protection: robust aggregation + detection of abnormal updates.

5. 3 AML/KYC

Objective: to reduce false-reject for thin-file without loss of sensitivity.
Approach: FL on document characteristics/payment patterns; PSI for sanctions/PEP lists; DP on units.

5. 4 Personalization/CRM

Objective: increase in LTV/retention without ethics violation and RG.
Approach: global preference model in FL + local adaptation of layers; exclusion of high-risk from "aggressive" offers; explainability for support.

6) Architectural scheme (reference)

1. Client silos: local phichepiplines (PII separated), local step training (E epochs).
2. Protection: DP-clipping/noise, channel encryption, Secure Agg keys.
3. Aggregator: TEE node with robust aggregator, deposit tracking, anomaly control.
4. Registers: Model Registry (versions, ε/ δ, thresholds), Feature Registry (feature policy).
5. CI/CD ML: fairness/privacy gates, poisoning tests, calibration and shadow runs.
6. Inference: centralized or co-inference with partners (MPC/TEE), journals without PII.

7) MLOps for FL

Policy-as-Code: white/gray/black lists of features, prohibition of proxy attributes; check at PR stage.
Pipeline hooks: group drift/calibration test, EO/EOp by segment, catching update anomalies.
Versioning: model/data/code + ε - accounting; "model cards" with sections Fairness & Privacy.
Catalog and line: connections "silo → aggregator → model version," "who and when taught," SLO freshness.
Observation: latency of FL rounds, proportion of participants, size/aggregation error, Attack AUC≈random.

8) Metrics and SLO

Quality: AUC/PR, calibration (Brier), uplift (for CRM).
Equity: EO/EOp deltas by country/channel/device.
Privacy: ε -usage, probability re-id, Attack-AUC (membership/inversion) ≈ 0. 5.
Reliability: participation of N participants ≥ the target threshold, the proportion of successful rounds, round time.
Security: percentage of abnormal updates rejected, poisoning incidents = 0.
Business: a decrease in chargeback/fraud, an improvement in RG outcomes, an increase in retention without an increase in disparity.

9) Templates (ready to use)

9. 1 FL Project Card

Task/Domain: (RG/AML/Payouts/CRM)

Topology: cross-silo/cross-device, aggregator hierarchy

Protection: Secure Agg, DP (ε/ δ), TEE/MPC, log policy

Participants: silo list, owners, trusted area

Metrics: quality, fairness, privacy, reliability, business KPI

Risks/mitigations: poisoning, non-IID, drift, jurisdictions

Release mode: shadow → canary → rollout, round frequency

9. 2 FL checklist before start-up

• Data contracts and feature policies agreed
• Secure Aggregation and Channel Encryption Configured
• DP parameters and ε accounting documented
• Robust aggregation and anomaly detection included
• Fairness thresholds/EO/EO and group calibration are set
• Shadow run passed, Attack-AUC ≈ random
• Incident plan (poisoning/privacy) and rollback ready

9. 3 Silo participation policy (fragment)

Minimum amount and quality of data to participate in the round

Mandatory local checks (DQ, calibration) before sending updates

Poisoning sanctions: exclusion/weight reduction/audit

Review of rights and logs: frequency and responsible

10) Implementation Roadmap

0-30 days (MVP)

1. Select 1 priority task (e.g. RG or Antifraud).
2. Define 3-5 silos, sign the policy and participation.
3. Deploy aggregator (TEE), enable Secure Agg and base DP.
4. Configure CI gates: fairness, privacy, poisoning tests.
5. Run 5-10 rounds of FL in shadow mode, compare with a centralized base.

30-90 days

1. Robust aggregators + anomaly detection, personalization by local layers.
2. Reduce traffic (quantization/deltas), enter partial participation.
3. Canary in sales for 5-10% of traffic, reports on SLO/ ε -usage.
4. Documents: FL project card, incident regulations, team training.

3-6 months

1. Extension to new silos/regions, hierarchical aggregation.
2. PSI/MPC for co-inference with PSP/vendors, private payment inference.
3. A single FL-observability dashboard, regular fairness/privacy audits.
4. Mass rollout, SLO and full coverage of high-impact tasks.

11) Anti-patterns

FL without Secure Aggregation/DP - "leaks through gradients."

Ignoring non-IID: one threshold/policy for all domains.
Lack of robust aggregation and monitoring of poisoning.
Logs with PII/feature dumps on the aggregator side.
"One-time trained and forgotten": without shadow/champion-challenger and review.

12) Connection with neighboring practices

Data Governance, Data Ethics, Confidential ML, Origin and Data Path, Reduce Bias, Model Monitoring, DSAR/Privacy - provide feature rules, transparency, metrics and managed releases.

Total

Federated Learning gives iGaming ecosystems joint intelligence without sharing raw data. With the right architecture (Secure Agg + DP + TEE/MPC), non-IID and poisoning resistance, and MLOps discipline, you get models that scale across markets and partners, withstand audits, and deliver consistent business value.