Skip to content

Federated and Privacy-Preserving Training Policy by Scale

Last updated: 2026-05-24

This page defines when an autonomy program should move from centralized training to federated, hybrid, or other privacy-preserving training patterns. Federated learning is not a maturity badge and not a default replacement for a governed central data pipeline. It is justified when raw data movement is legally, contractually, operationally, or economically constrained and when the model family can tolerate decentralized optimization.

The short rule: centralize consented and governed data when you can; federate or localize training when you must; never let privacy architecture bypass release evidence.


Decision Frame

Use federated or privacy-preserving training only when at least one hard constraint exists:

ConstraintExamplesBetter first option
Legal or contractual data residencyAirport, port, warehouse, customer, or country blocks raw data exportLocal data enclave plus central metadata
Competitive or tenant boundaryMultiple operators share platform but not raw operations dataTenant-isolated training exports
Raw-data volume or bandwidthMulti-LiDAR/camera logs cannot be uploaded in useful timeSelective upload and compressed event clips
Privacy riskFaces, bodies, badges, precise worker locations, cargo activity, or customer operationsRedaction, minimization, feature sharing
Institutional collaborationCross-company or cross-site model improvement is desiredCross-silo federation with governance
Local adaptationSites differ materially in objects, markings, weather, lighting, or operating rulesCentral base model plus local adapters

Do not federate just because the system has multiple vehicles. If the organization owns the data, can lawfully process it, can upload useful samples, and needs strong centralized evaluation, centralized training is simpler, more debuggable, and often better.


Scale Policy

MLOps scalePolicy stanceMinimum evidenceDo not do
S0 notebook researchSimulate federation only for learning or feasibilityClear note that clients are simulated and privacy is not provenClaim privacy benefit from a local split
S1 repeatable prototypeCompare centralized, local-only, and federated baselinesFixed federated partition, reproducible rounds, global and per-client metricsPick FL before measuring non-IID penalty
S2 single-product productionPrefer centralized or local fine-tuning unless data cannot moveData-use decision, privacy review, registry record, release packetLet federated round output move champion directly
S3 fleet and multi-siteUse hybrid training when sites differ or cannot share raw dataSite/client manifests, per-site holdouts, aggregation logs, privacy controlsOne global aggregate without local site approval
S4 regulated safety-criticalTreat FL output as candidate evidence, not approvalSafety-case links, privacy budget, poisoning/robustness review, immutable round evidenceApprove safety-critical behavior from aggregate metric only
S5 platform scaleProvide federation as a governed platform serviceTenant isolation, secure aggregation, policy-as-code, audit API, cost/SLO scorecardLet tenants define incompatible update schemas or privacy budgets

For autonomy, the strongest near-term production pattern is usually hybrid: central training on consented/owned data, site-local adaptation, and federated aggregation of small adapters or selected layers when cross-site raw-data movement is constrained.


Architecture Options

ArchitectureAdvantagesDisadvantagesBest fit
Centralized trainingBest observability, simpler debugging, strong global evaluationRaw-data movement, privacy, residency, bandwidth, tenant concernsOwned data, one site, consented centralized lake
Centralized with redaction/minimizationKeeps central tooling while reducing privacy riskRedaction QA burden; may remove useful cuesCamera-rich fleets and worker-visible sites
Site-local fine-tuningSimple governance, no cross-site optimizer complexityKnowledge does not transfer automaticallyOne customer/site with unique ODD
Hybrid central base plus local adaptersStrong base model and cheap local adaptationAdapter routing and compatibility must be governedMulti-site perception and semantic segmentation
Cross-silo federated learningRaw data stays local; shared model improves across sitesNon-IID drift, update security, orchestration overheadAirports, ports, warehouses, hospitals, enterprises
Cross-device federated learningUses many edge clientsClient churn, bandwidth, privacy attack surface, unreliable hardwareConsumer-scale mobile/vehicle fleets, rarely early AV programs
Split learningKeeps part of the model/data localLatency, activation leakage, complex deploymentSensitive features with stable connectivity
Secure enclave / confidential computeCentral compute with stronger isolationOperational complexity, hardware/cloud dependencyRegulated central training or secure aggregation services
Synthetic/embedding sharingAvoids raw-data transfer and may preserve task signalDomain gap, reconstruction risk, weaker safety evidenceRare scenes, scenario mining, low-risk analytics

Do not collapse these into one "privacy-preserving ML" bucket. Each option changes accuracy, observability, failure diagnosis, release evidence, and incident response differently.


Trigger Scorecard

Federation is justified when the scorecard shows a real constraint and the team can support the operational burden.

QuestionLow pressureMedium pressureHigh pressure
Raw-data export allowed?Yes, approvedAllowed with redaction or limited purposeBlocked by law, contract, tenant, or site policy
Data volume manageable?Upload samples and events easilyUpload budget constrains long-tail miningRaw logs cannot move in time or cost envelope
Site distribution differs?Similar ODD and objectsSome local terminology or weatherLocal classes, layouts, sensors, or rules differ strongly
Client count sufficient?One site or few vehicles2-3 sites with enough samples3+ meaningful clients or partner institutions
Evaluation available centrally?Strong central holdoutSome site holdoutsOnly local holdouts can legally hold raw labels
Privacy controls mature?Basic access controlRedaction and data inventoryPrivacy budget, secure aggregation, tenant isolation
Platform readiness?Scripts onlyPilot orchestration possibleRegistry, attestation, audit, monitoring, SLOs

Recommended decision:

Score patternDecision
Mostly low pressureStay centralized; improve data governance and selective upload
Privacy high, platform lowUse local fine-tuning or site-local models while building governance
Privacy/volume high, clients sufficient, platform mediumPilot hybrid central base plus federated adapters
Many clients, shared platform, strong governanceOffer governed cross-silo FL as S5 platform capability
Safety-critical artifact without local evidenceDo not promote; collect site-sliced evidence first

Required Contracts

ContractRequired fields
Client/site manifestSite ID, owner, jurisdiction, data residency, vehicle/sensor kit, ODD cell, allowed task
Local dataset manifestSnapshot ID, label schema, privacy tier, retention, allowed-use state, local holdout split
Training-round manifestRound ID, server model digest, client update digests, optimizer, local epochs, sample counts, aggregation algorithm
Privacy manifestSecure aggregation mode, differential privacy settings, privacy budget, clipping/noise policy, membership-inference test scope
Update attestationClient identity, code/container digest, local data manifest, produced update digest, policy result
Aggregation reportParticipating clients, dropped clients, weighting, robust aggregation, validation metrics, anomalies
Model registry recordBase model, federated/adapted artifact, client/site scope, global and local metrics, rollback target
Release packetClaim, evidence, ODD/site scope, local approval, safety-case link, monitoring and rollback plan

Federated updates are artifacts. They need digest-bound identity, policy checks, retention rules, and incident traceability.


Privacy and Security Controls

ControlWhat it helps withWhat it does not solve
Access control and local enclavesKeeps raw data at site or tenant boundaryGradient/update leakage
Secure aggregationServer sees only aggregate update, not individual updatesMalicious client updates, aggregate privacy leakage
Differential privacyQuantifies privacy loss by adding noise/clippingAccuracy loss, poor budget accounting, safety evidence gaps
Client update attestationProves client code and update identityData poisoning by approved client
Robust aggregationReduces outlier or malicious update impactSophisticated poisoning and non-IID bias
Trusted execution / confidential computeReduces infrastructure operator exposureModel/data leakage through outputs or poor policy
Redaction/minimizationReduces raw personal data before trainingDoes not prove privacy once features are learned
Privacy attack testingMembership/attribute inference and reconstruction probesCannot guarantee all attacks are absent

Privacy controls must be paired with utility evidence. A differentially private model that fails personnel detection, FOD recall, or false-free-space gates is not release-ready merely because privacy risk is lower.


Training Architecture Comparison

Model familyFederated fitPreferred patternCaveats
Small 2D/3D detectorsMedium to highFedAvg/FedProx or hybrid adaptersWatch class imbalance and site-specific labels
Large LiDAR/camera foundation backbonesLow for full-model FLCentral base plus LoRA/adaptersCommunication and optimizer stability dominate
Aggregated-map semantic segmentationMediumSite-local maps plus shared adapter or batch-level distillationSource-map acceptance and release-state labels stay local/evidence-bound
Open-vocabulary labelersMediumFederate prompt/eval metrics cautiously, not release labels directlyTaxonomy drift and prompt drift need governance
Planner or safety policyLowCentral/safety-reviewed training and local validationDo not federate safety authority without formal evidence
VLM/VLA assistantsMediumSite-local retrieval/prompt packs plus central eval packsPrivacy and hallucination risk; advisory only unless gated
Drift detectors and monitorsMedium to highFederated statistics with DP or secure aggregationMetrics can leak site activity if too granular

Federated learning is most attractive for representation adaptation, local perception robustness, and privacy-preserving analytics. It is least attractive for artifacts that require deterministic certification evidence or tightly controlled runtime behavior.


Evaluation Gates

GateRequired evidence
Centralized baselineSame architecture trained centrally on allowed data, when legally possible
Local-only baselinePer-site model/adaptor without aggregation
Federated baselineGlobal aggregate and per-client metrics under same splits
Non-IID stressSites/classes/weather/map states with skewed distributions
Privacy testMembership/attribute inference, reconstruction risk, privacy budget report
Robustness testPoisoned/stale/outlier client simulations and dropped-client rounds
Release testSite-sliced holdout, replay, shadow/canary, runtime package compatibility
Rollback testPrevious central/local/federated artifact remains executable

The release question is not "did the federated model improve the average?" It is "which sites improved, which regressed, which privacy promises hold, and which ODD cells are allowed to use this artifact?"


Airside and Non-Road Managed-Site Rules

  • Airport, port, campus, warehouse, and industrial customers may allow derived updates while blocking raw video/LiDAR export.
  • Local terminology matters. Prompt packs, class aliases, and label taxonomies should remain governed artifacts, not hidden per-client patches.
  • Site-local holdouts are mandatory when layout, lighting, markings, vehicle kits, aircraft/GSE mix, or pedestrian behavior differs.
  • Semantic-map training exports must preserve release-state labels locally. A federated adapter must not learn that stationary people, parked assets, FOD candidates, construction barriers, or ghost artifacts are permanent map truth.
  • Incident and safety/legal-hold data should not enter federation unless the legal/safety owner approves the purpose and retention state.
  • A customer-specific local adapter can be the right answer. Cross-site aggregation is not required when cross-site transfer creates more governance risk than value.

Operating Model

RoleResponsibility
Federation ownerAggregation service, round configuration, client onboarding, failure handling
Data/privacy ownerData-use approvals, residency, privacy budget, deletion propagation
Model ownerArchitecture, optimizer, evaluation, local/global performance tradeoff
Site ownerLocal holdout, ODD scope, operational acceptance, incident impact
Security ownerClient identity, secure aggregation, attestation, poisoning review
Release ownerRegistry alias, release packet, rollback, canary decision
Safety ownerSafety-case impact, waiver/expiry, hazard-slice release decision

At S5, the platform may own the federation service, but product and site teams still own release decisions.


Failure Modes

Failure modeConsequenceControl
Federation used before central baselineCannot tell if FL helped or hurtRequire central/local/federated comparison where lawful
Client data is too non-IIDGlobal model regresses local sitesPersonalized adapters and site-sliced release gates
Privacy budget ignoredClaimed DP protection is meaninglessPrivacy manifest and budget accounting
Secure aggregation treated as complete privacyAggregate still leaks or model memorizesDP, minimization, privacy attack tests
Malicious or broken client poisons updateGlobal model degradesClient attestation, anomaly detection, robust aggregation
Stale client updatesModel trains on outdated map/taxonomy/calibrationClient manifest freshness and update expiry
Site labels are incompatibleAggregation mixes different class meaningsTaxonomy compatibility gate
FL output moves directly to championRelease evidence is skippedRegistry and release packet gates
Local adapter routing is wrongVehicle loads wrong site modelCompatibility manifest and deployment scope checks
Deletion request cannot propagateDerived model/update remains contaminatedLineage graph and retraining/containment record

  • mlops-scale-research-scope.md - MLOps maturity ladder and research backlog.
  • federated-learning-fleet.md - detailed federated learning algorithms and airside fleet implementation notes.
  • mlops-migration-checklist-by-scale.md - transition gates before adding FL platform capabilities.
  • site-sliced-release-evidence-by-scale.md - local ODD-cell release approval.
  • model-governance-release-evidence.md - registry aliases, release packets, and rollback evidence.
  • secure-artifact-attestation-profile.md - update and model artifact attestation.
  • ../data-governance/fleet-data-privacy-governance.md - privacy governance and deletion propagation.
  • ../data-platform/fleet-data-pipeline.md - raw data movement, retention, and training exports.
  • continual-learning.md - federated continual learning and drift adaptation context.

Sources

Public research notes collected from public sources.