What Is Manipulated Media? Types, Digital Forgeries and How to Check

What Is Manipulated Media, and How Is It Different From a Deepfake?

Manipulated media is content whose meaning was changed after capture or generation. The test is the effect on the viewer: if an edit makes a recording show an event, a statement or a person differently from how it happened, the media is manipulated. Routine work such as colour correction, noise reduction or a crop that keeps the scene intact usually does not qualify.

A deepfake is the AI-built subset. The EU AI Act (Article 3(60)) defines a deep fake as AI-generated or manipulated image, audio or video that resembles real people, objects, places or events and would falsely appear authentic. Plenty of manipulated media involves no AI at all. The U.S. Department of Homeland Security's 2021 paper Increasing Threats of Deepfake Identities notes that audio tape could be cut and spliced to change a recording's meaning long before machine learning, and it describes cheapfakes as simple digital edits, such as slowed speech or accelerated video, that alter how an event is perceived.

The difference matters for the response. A cheapfake built from a genuine clip often has no synthesis traces for a detector to find, so the evidence that exposes it is the original footage. A face swap or cloned voice may carry generator traces even when the context looks plausible.

What Are the Main Types of Manipulated Media?

Manipulation techniques fall into a handful of families. Each changes a different part of the file, and each needs a different check.

Signal-level techniques alter the content itself: a swapped or reenacted face, a cloned or converted voice, an inpainted object, a composite of two photos, or audio stitched from several recordings. Presentation-level techniques leave the pixels mostly alone and change the frame around them: a clip trimmed to remove the answer to a question, footage slowed to make a speaker sound impaired, an old photo reposted with a new date, or a genuine file attached to a false claim.

What Is a Digital Forgery?

A digital forgery is a fabricated or altered digital artifact presented as genuine so that it misleads about its origin, content or meaning. The DHS Science and Technology Directorate's Digital Forgeries Report, dated 24 January 2023 and the first of five annual assessments required by the FY2021 National Defense Authorization Act, describes digital content forgery technologies as tools that fabricate or manipulate audio, visual or text content with the intent to mislead. It calls deepfakes a subset of digital forgeries created through machine learning.

The forgery frame is wider than media. A real photo can travel with fabricated EXIF data, a generated portrait can anchor a fake profile, and a genuine voice note can sit inside a forged chat history. The same report covers face morphing, where two faces are blended into one passport-style photo that both people can match, as an identity-document problem. Investigators therefore separate the file, the metadata, the account, the message thread and the claimed identity, and test each layer on its own. One authentic layer does not vouch for the others.

Intent is the other dividing line. Manipulated media describes what was done to the content. Calling something a forgery adds a claim about deception, which usually needs evidence beyond the file, such as who sent it and what they asked for.

How Is Manipulated Media Used for Fraud and Deception?

The most expensive documented case involved a fully fabricated video call. In January 2024 a finance employee at the engineering firm Arup in Hong Kong joined a call with what appeared to be the company's chief financial officer and other colleagues, all of them deepfake recreations, according to Hong Kong police. The employee made 15 transfers totalling HK$200 million, about US$25.6 million. Arup confirmed to CNN in May 2024 that fake voices and images were used.

The FBI's Internet Crime Complaint Center listed the broader pattern in a public service announcement on 3 December 2024: generated profile photos for romance and investment scams, forged driver's licences and banking credentials, fabricated disaster imagery for fake charity appeals, cloned voices of relatives in staged emergencies, and real-time video chats impersonating executives or police.

Human judgement does not close the gap. In a pre-registered experiment published in iScience in November 2021, Nils Köbis and colleagues found 210 participants identified deepfake videos with 57.6% accuracy, barely above chance. Participants labelled 67.4% of the videos authentic although half were fake, and the least accurate people were the most confident.

How Do You Check Whether Media Has Been Manipulated?

Work backwards from the circulating copy toward the original, and record each step so someone else can repeat it. The sequence below suits a newsroom, a fraud team or an HR investigation.

When you write up the result, name the change you can support, such as an inserted region, a splice at 00:41 or a caption that does not match the earliest upload. Leave the production method and the motive open unless separate evidence settles them. A report that says the available copy contains an altered region will survive scrutiny better than one that calls the whole item fake.

• Preserve the file as received, with its URL, sender, timestamps and any message that came with it, before it is deleted or re-encoded.
• Search for earlier copies with reverse image search or keyframe search and note the oldest reliable upload.
• Compare the circulating version with the source frame by frame or waveform by waveform to find cuts, speed changes or added material.
• Inspect provenance: C2PA Content Credentials where present, EXIF and container metadata, and platform upload records.
• Run forensic analysis on the best-quality copy you can obtain, and record the tool, model version and score.
• Corroborate the event or statement with records that did not arrive in the same message chain.

Where Does Automated Detection Help, and Where Does It Fall Short?

Forensic detectors earn their place on signal-level manipulation. Face swaps, generated faces, inpainted regions and synthetic speech leave statistical traces that software can measure even when a person sees nothing wrong. A good tool also localizes the evidence, which turns a vague suspicion into a specific region or segment an analyst can examine.

Detection has hard limits on the presentation layer. An unedited clip with a false caption is, at the pixel level, authentic, so no classifier will flag it; only the source record shows the lie. Re-encoding and resizing on messaging apps also erode the fine traces detectors depend on. Deepfake-Eval-2024 (arXiv 2503.02857), built from deepfakes that circulated in 2024, found open-source detectors lost about 50% of their AUC on video, 48% on audio and 45% on images compared with earlier academic benchmarks.

NIST's AI 100-4 report (November 2024) treats provenance tracking and synthetic content detection as complementary, and that is the practical lesson: run detection on the content, check provenance and source history for the context, and let a person make the final call when the stakes are high.

How Does Neural Defend Analyse Manipulated Media?

NeuroDetect is Neural Defend's forensic workspace for this job. An analyst uploads an authorized image, audio file, video or document into a case, and the analysis keeps the model and version context attached. Explainability heatmaps and attention views point to the regions, frames and audio segments that drove the result, which is the evidence an investigator needs to describe a manipulation precisely instead of labelling the whole file.

An LLM-assisted summary explains the findings in plain language and links each statement to the underlying evidence; the analyst still makes the final call. The case exports as a forensic report covering observations, uncertainty, provenance, model context and analyst notes, plus structured results for other systems.

Underneath is ARCAS 1B, Neural Defend's proprietary detection model. The ARCAS paper reports pooled accuracy of 99.99% on GenImage, 99.82% on AIGC-Detect and 97.65% on OpenFake. For real-time checks on faces in onboarding or account recovery, NeuroVerify runs six signals in one API call, including Face Manipulation Detection and AI Pixel-Level Manipulation, in under 200 ms p95 on eligible image requests. Neither product can tell you a caption is false; that still takes source work.