Deepfakes & Synthetic Media: understand, assess, respond.

The association’s role

The National Cybersecurity Association of Ukraine is a collaboration platform and an association of experts and market participants. In this area it develops methodological guidance, research and partnership initiatives against synthetic disinformation.

The foundation is public-private collaboration: an exchange of experience among public bodies, platforms, media, businesses and researchers, so that the verification approaches in circulation are transparent and understandable. There is no formal public-private partnership contract.

The association does not check individual items, run a lab or a detection service, or help remove content. No research report of its own exists yet: this page is an educational explainer with sources. An expert group for this area is only proposed.

What a deepfake is

A deepfake is synthetic audiovisual content or a photorealistic image made with artificial intelligence (AI). NIST notes that current multimodal models can generate this kind of content as well as disinformation text ¹.

The usual forms are face swaps, synthetic video, voice clones and generated images. Edited real media is a separate, lower-tech category: it needs no AI, yet it misleads just as well.

Why it matters: NIST observes that disinformation may erode trust in genuine evidence, and that generative AI can help create fraudulent content that impersonates other people ¹. ENISA reports the use of AI in voice phishing and impersonation fraud, including deepfakes ².

How to assess a suspicious video, voice message or call

A practical order is this: what is being asked of you, where it came from, what context it appeared in, whether there is independent confirmation, and only then tools. The first question often settles everything, because a scam needs your action, not your belief.

Visual “tells” are weak evidence. Ukraine’s Centre for Countering Disinformation notes that some AI-generated videos cannot be debunked without additional verification tools ³. The FBI lists subtle imperfections in an image as one cue ⁴, but treat it as a hint, not a test.

What detectors can and cannot tell you

In a benchmark of real deepfakes that circulated online in 2024, open-source detectors lost about 50% of AUC on video, 48% on audio and 45% on images compared with earlier benchmarks. Commercial and fine-tuned models did better but did not reach the level of forensic analysts ⁵.

WITNESS lists common problems with detection tools: false positives, difficulty across languages and variable accuracy. It advises being open about how you reached your conclusions ⁶.

The principle behind the association’s guidance: no detector can declare media “fake” or “real”. When a tool’s output is quoted, it should come with the tool and version, the date, the confidence level and the limits.

Provenance and labels

The C2PA coalition develops technical standards for certifying the source and history of media, known as provenance. Its specification site lists versions up to 2.4 and informative documents on security, harms and AI guidance ⁷. The technology is branded Content Credentials (C2PA).

Platforms are asking for labels too. TikTok requires labels on realistic content that is AI-generated or significantly edited ⁸. YouTube weighs whether synthetic content was disclosed when it assesses privacy complaints ⁹.

A label shows what the uploader or the platform declared. It is one signal, not a verdict. The same goes for provenance data: it helps when it is there, but it does not settle the question by itself.

If a fake is about you or your company: what platforms and authorities advise

1. 1Preserve: links, screenshots, date and time, account name. Delete nothing.
2. 2Do not spread the fake, even to debunk it.
3. 3Notify the platform.
4. 4Notify the authorities.
5. 5Warn your people: colleagues, family, partners.

YouTube takes privacy complaints about altered or synthetic likeness or voice, but the affected person must file them (with limited exceptions). You need the URL, timestamps and a description of what identifies you in the video ⁹. Telegram has a Report button for public content and abuse@telegram.org for takedown emails (include t.me links). For scammers impersonating you it points to @NoToScam, and it does not process private chats and groups ¹⁰.

For intimate imagery, including AI-generated images, adults can use StopNCII. A hash is created on your device, the image is not uploaded, and it works only with participating platforms ¹¹. AI-generated images fall within the service’s definition ¹². If the person was under 18 in the image, Take It Down from NCMEC is the route ¹³.

In Ukraine, the Cyber Police website has an incident form for phishing and malware, and a separate declaration page for other matters ¹⁴. An organisation can contact CERT-UA, which works with technical data and is not a punitive body ¹⁵.

Calls and voice clones

The FBI describes scams in which a cloned voice poses as a relative in trouble, and real-time AI video chats pose as an executive or official. Its advice: agree a secret word with close contacts, hang up and call back on a number you look up yourself, and limit how much of your voice and image is public ⁴.

The US Federal Trade Commission adds that a familiar voice is not proof, and that a request for wire transfers or gift cards is a red flag ¹⁶.

For journalists and fact-checkers

A basic sequence: the first source and first upload, metadata, context, independent corroboration, tool output with its limits, and only then a labelled conclusion. It rests on the advice to describe your process openly ⁶ and on the evidence that detectors are unstable ⁵.

The association does not publish private images or voices without a legal basis and consent. It runs no case library, lab or checking service: you cannot upload material here for verification.