Responsible Face AI
Trust, Privacy & Responsible Face AI
InsightFace supports enterprise teams building face AI products with privacy-aware deployment options, responsible use review, and commercial licensing workflows designed for sensitive AI applications.
Enterprise commitments
Bias mitigation as a product requirement
We treat demographic performance consistency as a core quality consideration and use evaluation findings to guide model and threshold review across relevant cohorts.
Training data governance
Training and evaluation data may include widely used public research datasets reviewed for applicable terms and licenses, licensed or commercially sourced datasets, appropriately consented private collections, and synthetic data.
Privacy-aware deployment
Our deployment options support biometric feature embeddings rather than raw-image retention where appropriate, including on-premise and offline processing in customer-controlled environments.
Algorithmic fairness
Evaluating demographic performance differences
We assess demographic performance where suitable evaluation data is available and use the results to inform model development and deployment guidance.
Cohort-aware evaluation: where datasets and labels support it, we evaluate false-match and false-non-match performance across relevant demographic cohorts.
Performance monitoring: internal evaluation is used to monitor and reduce cross-demographic differences in false-match and false-non-match performance.
Customer validation: customers should evaluate models against their intended population, capture conditions, operating thresholds, and use-case risks before production deployment.
Data ethics
Training data governance and sensitive-data handling
Data sourcing, intake review, identifier minimization, and deployment architecture are addressed as distinct parts of our governance approach.
Data intake and review: we maintain procedures designed to assess applicable usage rights, restrictions, and suitability before data is incorporated into model development. Our practices are designed with applicable privacy, biometric-data, and data-protection requirements in mind.
Identifier minimization: where appropriate, unnecessary metadata and direct identifiers are removed or minimized before data enters the model-development pipeline. Facial imagery and biometric-derived data are treated as sensitive data and protected accordingly.
Feature-based processing: deployment architecture can use biometric feature embeddings rather than retaining or transmitting raw images. Biometric embeddings are treated as sensitive biometric data and protected accordingly.
Dataset partnership inquiries
If you control or can document relevant rights for a dataset proposed for training or evaluation, contact our team. We may consider paid procurement after intake and suitability review.
contact@insightface.aiData Privacy & Retention
Data handling by deployment model
Data handling depends on the product, deployment model, configuration, and contract. Enterprise engagements can define retention, data flow, and operational controls appropriate to the selected architecture.
Customer Data Is Not Training Data
InsightFace does not use customer operational facial images, biometric templates, or inference data to train or improve general-purpose InsightFace models unless the customer has separately agreed to such use in writing.
For on-premise or offline deployments, operational biometric data remains within the customer-controlled environment and is not transmitted to InsightFace as part of normal inference.
Feature-based processing is available: face features can be extracted, compared, or stored as biometric feature embeddings rather than retaining raw images. Embeddings remain sensitive biometric data and require appropriate safeguards.
Cloud API retention and logging behavior is defined by the selected service and agreement. Where supported, customers may configure or contract for submitted images not to be retained beyond processing.
Retention periods and deletion practices for hosted deployments are documented in the applicable agreement and may be configured to support customer records and privacy requirements.
With on-premise and offline deployments, operational raw images, embeddings, and inference traffic can remain within customer-controlled infrastructure; the agreed architecture defines any external connectivity.
Deployment Options
Match the deployment model to your data flow
InsightFace supports multiple deployment patterns. Each one has a different data flow, and is appropriate for different sensitivity levels and operating constraints.
Cloud API
Hosted inference with no local model footprint.
- Data flow
- Images or pre-extracted embeddings are sent to a hosted endpoint over HTTPS; results are returned in the response.
- Fit for
- Pilots, centrally managed integrations, fast time-to-integration, and product teams that prefer a hosted model endpoint.
On-premise
Models run inside the customer's own infrastructure.
- Data flow
- Operational facial images, biometric embeddings, and inference data remain in customer-controlled infrastructure and are not transmitted to InsightFace as part of normal inference.
- Fit for
- Regulated industries, large-scale identity verification, KYC, access control, and customers with strict data residency requirements.
Mobile / Edge device
On-device inference via the InspireFace SDK.
- Data flow
- Images can be processed on the device; application design determines whether any application-level signals, such as a match decision, leave the device.
- Fit for
- Consumer mobile apps, embedded access control, IoT terminals, and offline / intermittent-network scenarios.
Responsible Face AI Use
Use-case review for sensitive face AI applications
Face recognition, face swap, and identity verification are sensitive applications. Commercial licensing and API access are reviewed against the intended deployment before activation.
We review the intended use case, deployment context, and end-user population before authorizing commercial face recognition or face swap models.
We expect customers to have a lawful basis for processing biometric data and to provide notice and choice to end users where required by local law.
We do not authorize use cases designed to enable unlawful or abusive surveillance, including surveillance that improperly targets vulnerable groups.
Authorization can be reviewed or revoked if usage materially diverges from the agreed scope.
Face Swap Usage Policy
Prohibited use cases for face swap models and APIs
Face swap models and APIs are licensed for reviewed creative, entertainment, and product use cases. The following are explicitly out of scope and will not be authorized.
Fraud, identity theft, KYC bypass, or any attempt to defeat identity verification systems.
Impersonation of real individuals without their explicit, verifiable consent.
Harassment, bullying, defamation, or content intended to humiliate or threaten a person.
Non-consensual sexual or intimate imagery of any individual, including synthetic NCII.
Unauthorized face replacement of private individuals, minors, or vulnerable populations.
Misleading political content, fabricated statements attributed to public figures, or content designed to manipulate elections or public discourse.
Enterprise Review Materials
Materials for security, legal, and procurement review
During procurement or due diligence, enterprise customers can contact our team to discuss or request the following materials. Availability and scope are confirmed for each engagement.
Training Data Governance Statement.
Customer Data & Model Training Policy.
Model / Algorithm Transparency Information.
Security and Deployment Architecture.
Responsible Use / Acceptable Use Information.
Evaluation Methodology and Performance Information.
Operational governance for enterprise buyers
Commercial discussions can address licensing scope, intended use, deployment architecture, data flows, and allocation of data-handling responsibilities.
We can provide product and deployment information to support customer-led assessments of applicable regional requirements, privacy controls, and internal approval workflows.
Evaluation and governance practices are reviewed as products, regulations, and deployment contexts evolve.
Need a trust or compliance review for your use case?
Talk with our team about model licensing, deployment architecture, data sourcing, and evaluation requirements for your market.
