GLOSSARY

Working definitions for a moving field.

We use these terms across the site in a specific way. Where a term has a primary source, a paper, a standard, or a statute, the citation is inline. Search to jump.

entries · last updated 2026-08-30

AAdversarial example: An input designed to cause a machine-learning model to misclassify, often imperceptibly different from a clean input. In synthetic-media forensics, adversarial examples are used both to attack detectors and to evaluate their robustness.
AASVspoof: A series of community challenges on automatic speaker verification anti-spoofing, including, since 2021, a deepfake-speech detection track. See Yamagishi et al., 2021.
AAudio deepfake: A synthetic audio recording, typically a voice clone of a target speaker, produced by a generative model trained on or conditioned on samples of that voice. Detection benchmark: ASVspoof. Detectors carry over poorly to voice generators they were not trained on, see when audio deepfake detectors fail on new voice generators.
BBEC (Business Email Compromise): A class of fraud in which the attacker impersonates a trusted contact, typically an executive, vendor, or finance counterparty, to authorize a wire transfer or change of banking details. The FBI's IC3 2024 Annual Report recorded $2.77B in U.S. BEC losses across 21,442 incidents. For the best-documented synthetic-media variant, see what the Arup deepfake actually proves.
CC2PA: Coalition for Content Provenance and Authenticity. An open technical standard for cryptographically signing the origin and edit history of digital content. The signed structure is called a Content Credential. Current spec: version 2.2 (1 May 2025). Products that have passed the C2PA conformance program are listed in a public registry, read in what the C2PA conformance registry actually lists.
CChallenge–response: An authentication pattern in which the verifier asks the asserted party for information only that party should know, or to perform an action only the real entity could perform. Useful against real-time deepfakes, see the Ferrari case (Bloomberg, 2024) and why deepfake executives fool people and what actually stops them, which explains why the check has to run outside the channel the caller used.
CCISA: U.S. Cybersecurity and Infrastructure Security Agency. Co-author of the September 2023 NSA/FBI/CISA Cybersecurity Information Sheet Contextualizing Deepfake Threats to Organizations (PDF).
CContent Credential: The signed manifest produced under the C2PA standard. Contains assertions about an asset's origin, edits, and use of AI tools, signed with X.509 certificates and embedded directly in the asset. It records what a signer asserted, not that the depicted events happened, see what Content Credentials establish, and what they do not.
CCross-generator generalization: The ability of a deepfake detector trained on outputs from one generator to maintain accuracy on outputs from a different generator. The central open problem in synthetic-media forensics. See Wang et al. (arXiv:1912.11035), and what a deepfake detector score does not tell you for the size of the reported drop.
DDeepfake: Colloquial term for synthetic media, typically face-swap or voice-clone content, produced by deep neural networks. A subset of the broader category of synthetic media.
DDEFIANCE Act: Pending federal bill, not enacted law. Disrupt Explicit Forged Images and Non-Consensual Edits Act. S.3696 (118th Cong.) passed Senate by unanimous consent on 23 July 2024; did not pass House. Reintroduced as S.1837 (119th Cong.); passed Senate again by unanimous consent in January 2026; still pending in House. Would, if enacted, create a federal civil cause of action for adults depicted in non-consensual intimate digital forgeries, liquidated damages of $150,000+; 10-year statute of limitations. Until it clears the House the federal regime here is criminal-only, as set out in what the TAKE IT DOWN Act changes.
DDFDC: Deepfake Detection Challenge dataset. 100,000+ paid-actor face-swap videos released by Meta AI in 2020 (Dolhansky et al.).
DDiffusion model: A class of generative model trained to invert a gradual noising process. Used in Stable Diffusion, DALL·E 3, and most modern open-weight image generators. Latent-space variant introduced by Rombach et al. (CVPR 2022). Detectors trained on GAN-family output often fail on diffusion-family output and the reverse, see why deepfake detectors fail on new generators.
FFaceForensics++: Benchmark dataset of 1,000 source video sequences manipulated with four pipelines (Deepfakes, Face2Face, FaceSwap, NeuralTextures). The defacto starting point for face-manipulation detection. Rössler et al., ICCV 2019. What a reported score on it does and does not mean is covered in what a deepfake detector score does not tell you.
FFakeCatcher: A real-time deepfake detection method by Ciftci, Demir, and Yin (IEEE TPAMI 2020) that exploits PPG signals. Biosignal features like these are one of the few threads that survive a change of generator, see why deepfake detectors fail on new generators.
FFrequency-domain analysis: A forensic method that converts an image into the frequency domain (e.g., via DCT or FFT) to expose upsampling artifacts characteristic of GAN-generated content. Frank et al., ICML 2020. Those artifacts are an upsampling fingerprint of one decoder family, which is why the method does not transfer, see why deepfake detectors fail on new generators.
GGAN (Generative Adversarial Network): A pair of neural networks, generator and discriminator, trained against each other. Goodfellow et al., NeurIPS 2014.
IIC3 (Internet Crime Complaint Center): The FBI's online portal for receiving cybercrime complaints from the public. Publishes the annual Internet Crime Report. Reporting URL: ic3.gov. Where the report fits in the first hour of an incident is set out in building a deepfake incident-response plan for smaller firms.
JJPEG Trust: ISO/IEC 21617. International standard for asserting media authenticity, provenance, attribution, IP, and integrity. Core Foundation (Part 1) approved at JPEG 105 in Berlin, October 2024 (press release).
LLiveness detection: Active or passive verification that the subject in front of a camera or microphone is a real, live person rather than a recording, photo, or synthetic stream. Used in identity verification and conferencing anti-fraud workflows.
MManifest (C2PA): The structured, signed JSON-LD claim set embedded in an asset under the C2PA standard. Contains assertions about origin, edits, and AI use, signed with an X.509 chain.
NNCII (Non-Consensual Intimate Imagery): Sexual or intimate imagery shared without the depicted person's consent, including AI-generated “digital forgeries” under the (enacted) TAKE IT DOWN Act and the (still-pending) DEFIANCE Act.
NNCMEC: National Center for Missing & Exploited Children. Operates the Take It Down hash service for those whose intimate imagery was taken when they were under 18.
PPPG (Photoplethysmography): The optical measurement of subtle skin-tone changes caused by blood circulation. Used as a deepfake forensic signal because synthetic faces do not preserve a coherent PPG signal. See FakeCatcher (TPAMI 2020).
PProvenance: The chain of facts about a digital asset's history, its source, the tools that created or altered it, the platforms that distributed it. Cryptographic provenance is the focus of the C2PA standard.
SStable Signature: A watermarking method that fine-tunes a latent-diffusion image generator to embed a recoverable binary signature in every output. Fernandez et al., ICCV 2023.
SStopNCII.org: A free service operated by SWGfL/Revenge Porn Helpline that hashes intimate images locally on a user's device and shares only the hash with participating platforms for proactive matching and removal. For adults 18+. stopncii.org.
SSynthetic media: Any image, audio, or video artifact produced or materially altered by a generative model, broader than “deepfake” and including consensual uses (film VFX, accessibility dubbing) and adversarial ones (impersonation, fraud, NCII).
SSynthID: Google DeepMind's family of watermarking systems for AI-generated text, image, audio, and video. SynthID-Image is documented in arXiv:2510.09263.
TTAKE IT DOWN Act: S.146 (119th Congress); signed into law 19 May 2025. Federal criminal prohibition on knowingly publishing non-consensual intimate visual depictions, with a 48-hour platform takedown obligation. Bill page. A practical reading for survivors, platforms, and counsel is in what the TAKE IT DOWN Act changes.
TTCPA: Telephone Consumer Protection Act. Federal statute regulating unsolicited calls, including (per FCC Declaratory Ruling 24-17) AI-generated voices in robocalls.
VVoice clone: A synthetic audio recording that imitates the voice of a specific target speaker, produced by a generative model trained on samples of that voice. Now within budget for ordinary fraud crews, see the Arup, Ferrari, LastPass, and 2019 Euler Hermes cases in the War Room, and deepfake wire fraud in real-estate closings for how the same technique lands on title and escrow workflows.
WWatermark (generative): A signal embedded in AI-generated content (visible or imperceptible) that allows downstream identification of the producing model. Examples: SynthID, Stable Signature. Limits: adversarial removal, perturbation/spoofing trade-offs (Saberi et al.). Why signing works only for cooperating actors is discussed in why deepfake detectors fail on new generators.

Continue

Definitions are upstream of accuracy.

References → The Research Lab →