ETSI Publishes Deepfake Safeguards TS 104 275: Prevention Starts at Design, Not Just Detection

What happened

The specification starts from the view that after-the-fact detection is not enough: ETSI said earlier deepfake efforts mainly focused on detecting manipulated content after it had been created, while the new specification treats prevention as a security and governance challenge to be handled throughout a system's lifecycle. It was developed by ETSI's Technical Committee on Securing Artificial Intelligence (TC SAI), builds on the EN 304 223 cybersecurity baseline for AI models and systems, and supports the objectives of the EU AI Act.

The measures listed in the specification include identifying possible misuse and setting safeguards before deployment, human review and intervention processes, content filtering, labelling of AI-generated content and user warnings, review of emerging threats, incident investigation and safeguard updates, and documentation and record-keeping. Alex Cadzow, rapporteur and secretary of ETSI TC SAI, said misuse of legal, commercially available software can also create harmful, misleading and illegal deepfake material, and that organizations need to address the risks starting from system design, with evidence that safeguards are in place.

The misuse cases covered include identity fraud and impersonation, as well as harmful deepfake material such as abusive images and fraudulent product sites. The document is a technical specification, not a legal compliance certificate: reporting notes that supporting EU AI Act objectives is not the same as acquiring legal force or certification, and applicable legal duties still need to be verified separately.

Key facts

1. Specification: ETSI TS 104 275, developed by the Technical Committee on Securing Artificial Intelligence (TC SAI)
2. Lifecycle stages: Secure design, development, deployment, maintenance and end of life
3. Core measures: Risk assessments, documentation, human oversight, misuse detection and response, content filtering, labelling and ongoing monitoring
4. Testing: Test scenarios and pass criteria to check whether safeguards are in place
5. Basis and aim: Builds on the EN 304 223 cybersecurity baseline and supports EU AI Act objectives
6. Risks covered: Identity fraud, impersonation, abusive images and fraudulent product sites, among other harmful deepfake material