Deepfake Glossary: Key Terms for Synthetic Media, Attacks and Detection

What Is a Deepfake Glossary of Key Terms?

This glossary collects the terms that come up when teams buy, deploy or investigate deepfake detection. Definitions are written for working use: each says what the thing is and, where it matters, how it differs from a neighbouring term. Terms are grouped by topic rather than alphabetically, because most readers arrive with a question about making, attacking, detecting or regulating synthetic media.

Where a standards body or regulator has defined a term, the definition here follows it. The core definitions are deepfake and synthetic media, and the two are not interchangeable.

• Deepfake. AI-generated or AI-manipulated image, audio or video that resembles real people, objects, places or events and would falsely appear authentic, following Article 3(60) of the EU AI Act.
• Synthetic media. Any content generated from scratch or substantially modified by AI, whether disclosed and harmless or deceptive. Deepfakes are a subset.
• Manipulated media. Content altered so that it misrepresents what happened, with or without AI.
• Cheapfake. Misleading media made with simple edits such as slowing, cutting or miscaptioning genuine footage, with no AI synthesis.
• Digital forgery. A fabricated or altered digital artifact, including documents and metadata, presented as genuine to mislead.

Which Terms Describe How Deepfakes Are Made?

Creation terms name the model families and techniques behind synthetic media. Knowing which one produced a fake helps explain which traces a detector may find.

• Generative adversarial network (GAN). Two neural networks trained against each other, one generating samples and one judging them, until the output is hard to tell from real data. Behind early face synthesis such as StyleGAN.
• Diffusion model. A model that learns to remove noise step by step and generates images, video or audio by starting from noise. Behind most current image and video generators.
• Autoencoder. A network that compresses a face into a compact code and reconstructs it; swapping decoders between two people produced the original face-swap deepfakes.
• Face swap. Replacing one person's face with another's in an image or video while keeping the original head movement and scene.
• Facial reenactment. Driving a real person's face with someone else's expressions and head pose, so the identity stays but the performance changes.
• Lip-sync. Regenerating only the mouth region so it matches new audio.
• Text-to-video. Generating an entire video clip, including people, from a written prompt.
• Voice cloning. Building a model of a specific person's voice from samples so it can say new words.
• Voice conversion. Transforming one speaker's live or recorded speech so it sounds like another person.
• Text-to-speech (TTS). Converting written text into synthetic speech, now often in a cloned voice.

Which Terms Describe Deepfake Attacks?

Attack terms describe how synthetic media reaches a target. In identity verification the delivery path matters as much as the fake itself, because each path defeats a different control.

• Presentation attack. Showing a fake to a camera or microphone: a printed photo, a mask, or a video played on a screen. ISO/IEC 30107-3 sets out how presentation attack detection is tested.
• Injection attack. Feeding fake video or images directly into an app's data stream, bypassing the physical camera, typically through a virtual camera, an emulator or a modified app.
• Virtual camera. Software that appears to the operating system as a webcam but outputs any video the attacker chooses, including a live face swap.
• Replay attack. Reusing a genuine recording of a person, captured earlier, to pass a check that expects live capture.
• Synthetic identity. A fabricated identity that combines real and invented details, often supported by a generated face and forged documents.
• Face morph. A single photo blended from two people's faces so that both can match it, used to obtain shared identity documents.
• Impersonation scam. Using a cloned voice or face of a known person, such as an executive or relative, to request money or access.
• Liar's dividend. The benefit a person gains by dismissing genuine evidence as a deepfake, a term from Robert Chesney and Danielle Citron's 2019 California Law Review article.

Which Terms Describe Deepfake Detection and Its Accuracy?

Detection terms are where buyers most often talk past vendors. A detector's headline number means little until you know which metric it is, on what data, and at what threshold. The public VendorBench-100 benchmark ranks 36 systems by MCC for that reason and warns that accuracy and F1 are unreliable on its imbalanced set.

• Deepfake detection. Software analysis that estimates whether media was generated or altered by AI and returns a score, ideally with evidence.
• Liveness detection (presentation attack detection, PAD). Checking whether a live person is physically present at the sensor. It does not by itself detect an injected deepfake.
• Artifact. A trace left by generation or editing, such as a blending edge, a frequency pattern or a spectral irregularity in speech.
• False positive. Genuine media flagged as fake. False negative: a fake passed as genuine.
• Threshold. The score at which a result triggers an action; moving it trades false positives against false negatives.
• MCC (Matthews correlation coefficient). A metric from -1 to +1 that uses all four cells of the confusion matrix and stays meaningful on imbalanced data.
• ROC-AUC. How well a detector ranks fakes above genuine files across all thresholds, from 0.5 (chance) to 1.0.
• Generalization. How well a detector performs on generators, devices and conditions it was not trained on.
• Explainability heatmap. A visual overlay showing which regions of an image or frame most influenced a detector's result.

Which Provenance and Regulatory Terms Should You Know?

Provenance terms describe information attached to content about where it came from. NIST's report AI 100-4 (November 2024) treats provenance data tracking and synthetic content detection as two separate families of digital content transparency techniques. Regulatory terms set what organizations must disclose or do.

• C2PA. The Coalition for Content Provenance and Authenticity, which publishes an open technical standard for recording the origin and edit history of digital content.
• Content Credentials. The consumer-facing name for C2PA provenance information, likened by the coalition to a nutrition label for digital content.
• Digital watermark. A signal embedded in content, visible or invisible, that marks it as generated or identifies its source; Google DeepMind's SynthID is one example.
• Content fingerprinting. Computing a compact signature of content so copies can be matched later, even after minor changes.
• Metadata. Data stored with a file, such as camera model, software and timestamps; easy to strip or forge.
• Chain of custody. A documented record of who collected, handled and stored evidence, and when.
• EU AI Act Article 50. Transparency duties applying from 2 August 2026, including machine-readable marking of AI outputs and disclosure of deepfakes.

What Do Neural Defend's Product Terms Mean?

Neural Defend's own vocabulary maps onto the terms above. NeuroVerify is the real-time API: one request checks a face image, video or audio and returns a unified risk score from six signals (Visual Safety Screening, Live Presence Verification, Face Manipulation Detection, AI Generative Model Analysis, AI Pixel-Level Manipulation and AI Content Compliance), in under 200 ms p95 on eligible image requests. NeuroDetect is the forensic workspace, where cases hold authorized media, explainability heatmaps and an exportable forensic report. NeuroCall (Beta) covers voice calls.

ARCAS 1B is Neural Defend's proprietary deepfake detection foundation model. On VendorBench-100, the public still-image benchmark ranked by MCC, Neural Defend placed first of 36 systems with MCC 0.876 against 0.646 for the next vendor; one benchmark author also co-wrote the ARCAS paper, and the data is public for anyone to rerun.