1-3 | Scope, references and terms | 0 |
4.1.1 | Training data source selection: sample before and after collection, reject sources above five per cent illegal or unhealthy content | 0 |
4.1.2 | Diversity and mix of training data sources | 0 |
4.1.3 | Management and traceability of training data sources: licences, collection records, commercial contracts and user authorisation | 0 |
4.2.1 | Filter all training data of every type before training | 0 |
4.2.2 | Intellectual property protection in training data and generated content | 0 |
4.2.3 | Personal information in training data only with consent or another lawful basis | 0 |
4.3.1 | Annotation personnel: training, assessment, qualification and separation of duties | 0 |
4.3.2 | Annotation rules for functional and security annotation | 0 |
4.3.3 | Accuracy of annotated content: batch sampling, re-annotation and review | 0 |
4.3.4 | Isolated storage of security annotation data | 0 |
5.1 | Model training security: content security as a core optimisation metric, framework audits and backdoor detection | 0 |
5.2 | Model output security: ninety per cent pass rate, accuracy, reliability, refusal behaviour and content labelling | 0 |
5.3 | Model monitoring and evaluation: input monitoring, routine testing and emergency handling | 0 |
5.4 | Model update and upgrade security: policy and self-assessment after major changes | 0 |
5.5 | Isolation of training and inference environments | 0 |
6.1 | Service suitability: justify use cases, protect high-risk scenarios and protect minors | 0 |
6.2 | Service transparency: publish audience, scenarios, purposes, limitations, model summary and personal information use | 0 |
6.3 | User input used for training only with a prominent notice and an easy opt-out | 0 |
6.4 | Public and user complaint and reporting channels with rules and time limits | 0 |
6.5 | Serving users: input detection with published suspension rules, and adequate monitoring staff | 0 |
6.6 | Service stability and continuity: backup and recovery | 0 |
6.7 | On-device models: activation, on-device security module, keyword updates and model patching | 0 |
A | Appendix A: the 31 security risks in training data and generated content | 0 |
B | Appendix B: security assessment reference method | 0 |