Computer vision in iGaming

1) Why CV pipelines iGaming platform

KYC/AML: OCR documents, authentication, liveness/anti-spoofing.
Antifraud/risk: detection of bots/multi-accounts (behavioral + visual), identification of "screen sharing" and proxy devices.
Marketing/ASO: moderation of creatives (text/symbols/rating 18 +), brand safety, A/B visual elements.
Operations/QA: UI automatic regression tests, visual telemetry of lags/crushes.
Streams/social networks: extracting events, logos, games/providers, tonality and violations.
Responsible Gaming: control of visual communications (lack of aggressive patterns for vulnerable groups).

2) Key scenarios and solutions

2. 1 KYC: document + person

OCR: extraction of full name/date/document number, format validation, comparison with the application.
Face match: Comparing selfies to photos in a document.
Liveness: passive signs (micro-motion, Moiré, blink) and active (prompt-challenge).
Document authenticity: watermarks/fonts/microprint, photoshop detection.

2. 2 Antifraud and safety

Device cam check (where allowed): signs of playback from the screen/mask.
Multi-account: combining CV signals (selfies/backgrounds) with behavioral and device graphs.
Content policies: blocking payment card/passport images in open channels.

2. 3 Marketing/Creative/ASO

Moderation: detection of prohibited symbols/slogans, "18 +," QR/links, bets.
Brand safety: compliance with guides by logo, colors, location.
A/B: automatic composition analysis (CTA, contrast, "workload"), correlation with CTR/CR.

2. 4 Streams & Videos (Games/eSports/Influencers)

Logo/Game detection: providers' promo counters.
Highlight mining: clips by event (big win/bug/connection break).
Video moderation: P-rating, gambling content by hour of display/jurisdiction.

2. 5 UI/QA

Visual regression: comparison of screenshots by page/version/device.
Optical telemetry: frame timings, render omissions, "blinking" elements.
Accessibility: checking contrast/size/alt-text in creatives and pages.

3) Architectures and deployment

On-device (mobile SDK, WebAssembly): instant liveness/OCR without sending frames (privacy by default).
Edge (PoR/region): low latency and geo-isolation of data/keys.
Cloud: heavy models (detection, segmentation, video analysis), asynchronous tasks.
Confidential inference: TEE/SGX for VIP/payout; protected pipelines.
Hybrid: Easy on-device pre-validation → accurate edge/cloud validation.

4) Data and augmentation

Collection: consent, PII disguise, geo-retention policies.
Synthetics: document/selfie generation with lighting/angle/noise variations; domain randomization.
Augmentation: blur, motion, glare, print-scan, screen-to-screen (screen re-capture), JPEG artifacts.
Balance: classes "spoof," "photo from the screen," "mask," "multi-exposure" - at least positive.
Markup: active learning; QA-double verification of disputed cases.

5) Models and patterns

Classification/detection: YOLOv8/YOLOv9, EfficientDet, ViT/DETR; for logos - specialized detectors.
Segmentation: SegFormer/Mask2Former (background/masks, path document).
OCR: TrOCR/ABINet/CRNN + rectification; multilingual support.
Face: ArcFace/FaceNet for embeddings; Anti-spoof CNN/ViT; liveness by micro-movements.
Video: SlowFast/X3D/TimeSformer; for highlights - event classifiers + Energy-based filters.
Multimodality: CLIP-like models for creatives (image + text).

6) Pipelines (end-to-end view)

6. 1 KYC/Liveness (edge + cloud)

1. On-device: qualifier frame (sharpness/lighting) → passive liveness.
2. Edge: OCR of the document, comparison of face-embeddings, spoof-check; risk-rate.
3. Cloud: manual verification of disputed cases (HITL), audit, DSAR log.

6. 2 Moderation of creatives

1. Ingest creatives (from DAM/admin panel) →

2. Detection of text/symbols/logos →

3. Classification of "allow/flag/deny" by jurisdiction →

4. API to ad engine + reporting.

6. 3 Visual regression UI

1. Script/screenshot generator by device/local →

2. Per-pixel/per-object comparison + tolerances →

3. Alert in PR/CI; auto-capture before/after.

7) Quality and SLO metrics

Optional: Bias/Fairness by skin/lighting/camera; Privacy (zero PII frame/log leaks).

8) Security, privacy and compliance

Biometrics-by-design: minimization/locality (on-device), encryption, shelf life by policy.
Tokenization of face embeddings, prohibition of reversibility, separate keys.
DSAR/delete: search by subject token, crypto erase.
Legal Hold: Video/footage freeze for investigations.
Jurisdictions: geo-isolation of data/keys, different 18 +/advertising rules.
Audit: immutable inference/decision logs (WORM), explainability of boundary cases.
Tricks of intruders: protection against re-capture, adversarial patterns, rate limiting.

9) Observability and alerts

Online metrics: latency p50/95/99, error rate, saturations (GPU/CPU/IO).
Quality: drift by lighting/cameras/countries; growth of APCER or FPR.
Operating system: queue of controversial cases, SLA manual verification.
Alerts: surge in deny misses/false positives, drop in OCR accuracy.

10) Integrations (API/Contracts)

10. 1 KYC Service

yaml api: /v1/kyc/check request:
selfie: image_token document_front: image_token document_back: image_token country: "EE"
purpose: "account_opening"
response:
scores: {face_match: 0.93, spoof: 0.02}
ocr: {name: "IVAN IVANOV", dob: "1994-02-14"}
decision: "allow    manual    deny"
trace_id: "..."
privacy: {pii: true, tokenized: true}

10. 2 Moderation of creatives

yaml api: /v1/creative/moderate request: {image_token: "...", market: "TR", channel: "display"}
response:
violations: ["age_rating_missing","prohibited_text"]
decision: "deny"
trace_id: "..."

11) MLOps for CV

Registry: model/data/augmentation/versions; limitations of use.
Releases: shadow/canary/blue-green, rollback by FPR/latency.
Tests: golden set with "heavy" cases (masks, glare plastic, reshoot screen).
Monitoring: drift light-feature (illumination, sharpness), bias-reports.
Cost: INT8/FP16, sparsity, batch-size, preprocessing cache, light/heavy routing model.

12) Templates (ready to use)

12. 1 Inference Policy (SLO/Privacy)

yaml cv_service: vision.core slo:
p95_latency_ms: 300 success_rate: 0.995 privacy:
store_frames: false biometrics_tokenized: true retention: "P30D"
monitoring:
spoof_apcer_max: 0.03 ocr_cer_max: 0.06 bias_gap_pp_max: 3

12. 2 KYC module start-up checklist

• On-device pre-validation and passive liveness enabled
• CER/WER on ≤ threshold golden set
• Bias report on cameras/lighting/document types
• Shadow 5-10% of applications, manual revision of disputed
• DSAR/removal and Legal Hold verified
• APCER/BPCER and latency alerts

12. 3 Runbook "APCER Growth"

1. Check dashboard by cameras/countries; Define hot segments.
2. Switch to the "heavy" anti-spoof model on Edge in these segments.
3. Tighten the thresholds, enable the active check (blink/prompt).
4. Update augmentation and golden set; post-mortem.

13) Implementation Roadmap

0-30 days (MVP)

1. KYC: OCR + basic face-match, passive liveness on-device, manual verification of controversial.
2. Moderation of creatives: rules + text/logo detector; deny list by jurisdiction.
3. UI-regression: visa-snapshots of top screens, PR-gate by diff%.

30-90 days

1. Anti-spoof ViT, active promptas; document/selfie synthetics.
2. Video analytics of streams: logo/highlights; reports to providers.
3. bias/fairness reports, drift monitoring; canary releases, SLO alerts.

3-6 months

1. Confidential inference (TEE) for VIP/payout.
2. Full control of brand safety and A/B creatives with correlation to CR/ARPPU.
3. Auto-generation of golden sets from controversial cases; champion-challenger configs.
4. External integrations with providers/CUS partners for signed webhooks.

14) Anti-patterns

The lack of a golden set and bias audit → "good on average, bad on the edges."

Storage of "raw" personnel without need and time; logs with PII.
Liveness is only active (no passive) or vice versa.
Universal thresholds for all countries/cameras/scenes (ignore seasonality/illumination).
Run heavy models without profiling and latency/cost budgets.
Moderating creatives with the "last step" before release is expensive and late.

15) Related Sections

KYC/AML and access control, DataOps practices, MLOps: model exploitation, analytics and metrics APIs, Feedback sentiment analysis, Data stream alerts, Data ethics and transparency, Data retention policies.

Result

Computer vision is not a "separate neural network," but part of the production pipeline of data and risks: from on-device privacy and geo-isolation to MLOps and quality alerts. The correct CV architecture reduces fraud and manual checks, speeds up KYC, makes marketing safe and measurable, and the product more stable and affordable.