Skip to content

InVID-WeVerify Verification Plugin

FreeMulti-tier

Vendor: invid-project.eu | Type: Detector + non-detector | Cost: Free

TL;DR

A free Chromium browser extension that bundles reverse-image search, video keyframe extraction, EXIF metadata, OCR, image forensics, deepfake screening, and Hiya-backed voice-clone detection into one plugin used by AFP, Bellingcat, VERA Files, and trained fact-checkers across Southeast Asia. The hub of Pillar 1B.

What it does

InVID-WeVerify is a single browser extension that turns a fact-checker's desk session into a multi-tool forensic pipeline. It accepts URLs, images, and videos, then returns: keyframes pulled at scene-change boundaries; reverse-image search results across Google, Bing, Yandex, Baidu, and TinEye; EXIF and metadata extraction; OCR on text inside images; basic image-forensics outputs (ELA, copy-move detection); a deepfake-screening tab that uploads frames to CERTH servers for an "experimental" probability score; and a voice-clone module that wraps Hiya for audio classification. It also captures WACZ archives of source URLs for chain-of-custody preservation.

The plugin is the workhorse of Pillar 1B because its non-detector modules (reverse-image, archived versions, metadata) usually resolve the question without the user ever invoking the deepfake tab. Where the case escalates, the deepfake and voice-clone modules give a cautious-detector signal that combines per Anchor 3 as one signal class.

When to use it

  • A WhatsApp-forwarded image or video lands in your queue and you need to check whether it has appeared elsewhere online before scoring it for AI generation.
  • A Tagalog or Bahasa video circulating on Facebook needs keyframe extraction and reverse-image cross-checking before you decide whether to escalate to enterprise detectors.
  • You need WACZ-format archival preservation of a source URL before it is removed or edited, with the archive capturing live state plus metadata.
  • You are running a workshop and want one extension that demonstrates the full Pillar 1B verification stack without installing five separate tools.

Independent accuracy

Vendor claim vs independent assessment

Vendor claim (deepfake tab): the deepfake-screening tab is documented in the plugin as "experimental" with reported accuracy in the 70-90% range that degrades on newer generation methods.

Vendor claim (voice-clone module / Hiya): Hiya's marketing, quoted in the DW Innovation audit, promised "free, real-time detection, multi-language... detects voices created by all popular voice synthesis tools" but did not guarantee complete accuracy.

Independent finding (Hiya audio module): the DW Innovation civil-society audit (Synthetic Audio Detectors Put to the Test, September 2025) tested Hiya on a 10-sample multilingual dataset and found it correctly identified 4 of 10 cases, misidentified 3 of 10, and returned 3 of 10 inconclusive. The audit also noted contradictory visualisations within the interface.

No independent SEA-specific benchmark identified as of May 2026 for the deepfake tab on tonal-language video or SEA-platform-compressed footage. The audit's 10-sample test covered multiple languages but published no per-language or Asian-language breakdown.

Limitations

Limitations

  • Deepfake tab is documented as "experimental"; reported accuracy degrades on new generation methods.
  • Frames uploaded to the deepfake tab go to CERTH servers and are retained for 30 days; a privacy concern for source-identifying content.
  • Twitter/X features have been broken since the API change.
  • UI is English by default; localisation covers French, Spanish, Greek, Arabic, Italian, German, Japanese, and Hungarian, but none of the seven SEA languages (Bahasa Indonesia, Lao, Malay, Filipino/Tagalog, Sinhala, Tamil, Thai). Trained fact-checkers treat the English UI as workable; the gap matters for semi-trained volunteers.
  • Voice-clone module wraps Hiya; carries the DW Innovation cross-language unreliability finding verbatim.

Privacy and threat model

The plugin's non-detector modules (reverse-image, EXIF, OCR, archiving) operate without uploading the source file to InVID-controlled servers; they query third-party search engines or run locally in the browser. The deepfake tab is the exception: it uploads video frames to CERTH (Centre for Research and Technology Hellas, the Greek research partner in the vera.ai consortium) and retains them for 30 days. The voice-clone module routes audio to Hiya, a US vendor, under that vendor's retention policy.

For Sri Lanka, Lao, and other surveillance-environment work, classify the source file before invoking the deepfake or voice-clone tabs. If the file is source-identifying (a frame containing a recognisable informant, an audio clip with diagnostic background noise), the upload itself is the risk, not the verdict that comes back.

Source-protection override (S1 — source-identifying upload risk)

Uploading source-identifying frames or audio to the deepfake or voice-clone tabs exposes the source to a third-party server with a 30-day retention window (CERTH) or a US-vendor jurisdiction (Hiya). Mitigation steps:

  1. Classify the file as public, sensitive, or source-identifying before any upload.
  2. For source-identifying material, do not use the deepfake or voice-clone tabs. Use the reverse-image, EXIF, OCR, and archiving modules instead; these do not upload the source file to InVID-controlled servers.
  3. If a deepfake-tab verdict is operationally necessary, strip identifying context (crop, blur background, redact audio) on the local machine before uploading the redacted version.
  4. If you uploaded by mistake, request CERTH deletion via the vera.ai contact channel and note the disclosure to the source.

Source-protection override (S9 — cross-border data transfer)

The deepfake tab routes uploads to CERTH (EU jurisdiction) with a documented 30-day retention window; the voice-clone tab routes to Hiya (US jurisdiction) under that vendor's terms. This is the toolkit's reference case for cross-border vendor retention. Mitigation steps:

  1. Treat the 30-day CERTH window and the Hiya US-vendor regime as separate compliance facts; both apply when both tabs are used on the same case.
  2. Strip identifying context before any deepfake-tab or voice-clone-tab upload, then request CERTH or vendor deletion at case close via the documented support channel.
  3. The non-deepfake modules (reverse-image, EXIF, OCR, WACZ archiving) do not transit the source file to InVID-controlled servers; route source-identifying work through these instead.

Country and platform applicability

  • Indonesia: documented use through training and Tempo workflow; pairs well with Kalimasada at the WhatsApp-intake side.
  • Laos: plugin operates language-agnostically on visual modules; UI is English. Honest gap: deepfake and voice-clone tabs are not benchmarked on Lao-language audio or on Lao-region content.
  • Malaysia: general OSINT use; no documented MAFINDO or Sebenarnya pipeline integration recorded .
  • Philippines: documented heavy reliance at VERA Files (Nieman Lab quote); standard for AFP Fact Check Philippines and the #FactsFirstPH coalition (regional case documentation).
  • Sri Lanka: language-agnostic visual modules apply; Watchdog and Hashtag Generation use the plugin alongside their own Sinhala/Tamil OSINT stack. Honest gap: no SL-specific benchmark for the deepfake or voice-clone tabs.
  • Thailand: AFP training at Chulalongkorn University, August 2025; standard for Thai PBS workflow alongside SynthID Detector (regional case documentation).

Platform applicability: works on any web-accessible image or video, relevant across Facebook, Facebook Groups, TikTok, YouTube, Telegram. The X/-x modules are broken since the API change.

How to access

Free download from the Chrome Web Store, Microsoft Edge Add-ons, or the Opera add-on store (search "InVID-WeVerify" or "vera.ai Verification Plugin"). The repository at AFP-Medialab/verification-plugin on GitHub publishes source and release notes; v0.87 shipped in July 2025; commits continue through April 2026. Account creation is not required for core modules; some advanced features (Hiya audio analysis, DBKF queries) require registration.

Cost (current as of 2026-05)

Free. The plugin is open-source under MIT licence, funded through the vera.ai consortium under EU Horizon Grant Agreement 101070093. No paid tier; no quota for ordinary use. Vendor wraps inside the plugin (Hiya voice-clone module) operate under those vendors' separate cost structures, but the plugin itself imposes no fee.

Quickstart

  1. Install the extension from your browser's add-on store.
  2. Right-click any image or video on a webpage and choose the "InVID-WeVerify" submenu.
  3. For an image, run "Image search" first; this fans out reverse-image queries to Google, Bing, Yandex, Baidu, and TinEye in one click.
  4. For a video, choose "Video → Analysis" to extract keyframes, then right-click each keyframe to reverse-image-search it.
  5. Use the "Image forensics" tab for ELA and copy-move detection (browser-local; no upload).
  6. For metadata, choose "Image → Metadata"; this reads EXIF in the browser without uploading.
  7. Only if non-detector modules are exhausted and the case warrants it, classify your file per the danger admonition above and then try the "Deepfake" or "Audio analysis" tabs; these upload to CERTH or Hiya respectively.
  8. Capture a WACZ archive of the source URL via "Archive" before the page changes.

In the toolkit's workflow

Multi-pillar tool: serves provenance (metadata module), source-history (reverse-image, archiving), and cautious-detector (deepfake and voice-clone tabs) per Architectural Anchor 1. Sits at the centre of 1B.1 Professional Verification, with cross-cell roles in 1A.2 First-Line Triage video and 2A.2 reverse-image / geolocation.

Standard combinations:

  • With ExifTool for command-line metadata work on the same source file when batch processing is needed.
  • With Sherloq when the source file cannot leave the user's machine (Sri Lanka, Lao surveillance contexts).
  • With Hive AI at 1A.1 / 1A.3 for a parallel detector signal, counted as one detector signal class with the InVID deepfake tab per Anchor 3.
  • With Auto Archiver at 2A.2 for evidence preservation beyond the WACZ module's capture.
  • With Meedan Check at 2B.1 when the same content is also being processed through a tipline.

Decision-tree references: T1 image triage routes to InVID at T1.5 (reverse-image search) and T1.6 (metadata) as the default first PV step. T2 video triage routes to InVID's keyframe extraction at T2.4 and reverse search at T2.5. T6 source-protection routes the deepfake tab through the S1 sub-routine before any upload.

Conflict resolution behaviour

When the deepfake tab's verdict disagrees with another detector (Hive, Sensity, Deepware), the InVID tab carries the lower evidence weight in most cases because (a) the underlying CERTH model is documented as experimental, with vendor accuracy in the 70-90% range, and (b) it has not been independently audited at the level of the Deepfake-Eval-2024 benchmark or the DW Innovation audit. When it disagrees with a non-detector signal (reverse-image hit, provenance manifest, archived earlier instance), the non-detector signal wins because Anchor 2 mandates two non-detector signals before any strong public claim. The InVID deepfake verdict alone is never sufficient for a publishable claim.

Override notes

Override notes

  • Multi-pillar declaration: the plugin serves provenance, source-history, and cautious-detector pillars; the deepfake and voice-clone tabs are explicitly framed as cautious-detector with a weak-signal caveat.

  • Deepfake-tab caveat (c4-deepfake-tab-caveat): the deepfake tab is "experimental" and accuracy degrades on new generation methods. Use only after non-detector modules are exhausted and treat the verdict as one weak signal.

  • CERTH 30-day retention mitigation (d4-mitigation-CERTH-30day): the deepfake tab uploads to CERTH for 30 days. Mitigation is bound into the danger admonition above (S1 source-protection sub-routine).

  • Safety-before-upload binding (p2-safety-before-upload-binding): pre-upload sensitivity classification is required for the deepfake and voice-clone tabs. See the danger admonition above.

Sources