Model security
Measures to protect models from theft, tampering, or misuse.
How the laws define this
Both instruments establish that model security consists of technical and organizational measures to protect the confidentiality, integrity, and availability of AI models against misuse. They explicitly agree on core mechanisms, specifically mandating access controls and provenance tracking. However, they diverge on asset scope and specific threat controls. The Guidelines for Secure Adoption and Use of Artificial Intelligence (National Cyber Security Agency) broadens coverage to include related assets alongside models, emphasizing anti-poisoning measures, adversarial robustness testing, and protections against model extraction. In contrast, the Buku Putih Peta Jalan Kecerdasan Artifisial Nasional (White Paper / Roadmap for National AI) limits its focus to protecting the model itself from theft, tampering, and adversarial exploitation, while distinctly adding incident response protocols to its scope.
Synthesised from the 2 statutory definitions below. Each one is quoted in full, with its source.
Definitions (2)
Technical and organizational measures designed to protect model integrity, confidentiality, and availability against theft, tampering, adversarial exploitation, and misuse, including access controls, provenance tracking, and incident response protocols.
Technical and organisational controls to protect the confidentiality, integrity and availability of AI models and related assets, including anti‑poisoning measures, access controls, provenance tracking, adversarial robustness testing and protections against model extraction or misuse.
Related Terms
Secure AI
Protection against malicious manipulation and unauthorized access....
Security (AI)
Protection of AI systems from malicious attacks....
Trust/security
Privacy, data governance, robustness and cybersecurity....
model transparency
Clarity about a model's operation, inputs, outputs and limitations....
Model robustness
Resistance of AI models to errors, attacks and unexpected inputs....