Model and data protection
In one line
The model and the training data are the crown jewels - protect their confidentiality and integrity across the whole lifecycle, not just the network around them.
The threats
- Model theft / exfiltration - the trained model is valuable IP; protect against extraction.
- Training-data confidentiality - datasets often contain sensitive/regulated data.
- Data poisoning (integrity) - corrupting training data to corrupt the model's behaviour.
- Supply-chain risk - third-party pretrained models and datasets may be tampered or tainted.
Controls
- Encryption at rest and in transit (Transport security and encrypted-traffic implications) for models and datasets.
- Strict access control to model/data stores (Storage and checkpointing for AI).
- Integrity, lineage, and assurance (Regulations, governance, and AI security policy) - know where data/models came from and that they're unaltered.
- Provenance checks on third-party models/datasets.
- Apply the CIA triad to AI assets (Security design fundamentals - CIA and defense in depth).
Recommendation / justification
Treat models and datasets as top-tier assets: encrypt, tightly control access, verify integrity/provenance, and govern lineage. Justify by the fact that the value and risk concentrate in the model and data, which is distinct from (and additional to) securing the fabric.
What would change this (mid-scenario twist)
- Use of external/pretrained models -> provenance and supply-chain controls.
- Regulated training data -> encryption + strict access + auditable lineage.
Validation checks
- Are models and datasets encrypted and access-controlled, with integrity/lineage tracked?
- Is third-party model/dataset provenance verified (supply chain)?
IPv6 / dual-stack note
Data/model protection is at the storage/app layer (address-agnostic); ensure transit encryption covers both families.
Related
- Transport security and encrypted-traffic implications
- Regulations, governance, and AI security policy
- Security design fundamentals - CIA and defense in depth
- Storage and checkpointing for AI
Spaced repetition
Name the main model/data threats in AI.
Model theft/exfiltration, training-data confidentiality, data poisoning (integrity), and supply-chain risk (tainted third-party models/datasets).
Data poisoning attacks [...].
Data poisoning attacks integrity - corrupting training data to corrupt the resulting model's behaviour.
Model/data protection rests on [...] (CIA applied to AI assets).
Model/data protection rests on encryption + strict access control + integrity/lineage/provenance (CIA applied to AI assets).
Sources
- Cisco AI/ML blueprint; MITRE ATLAS / responsible-AI security references.