Deepware Scanner¶
Vendor: deepware.ai | Type: Detector | Cost: Freemium
TL;DR
A press-button web and mobile video deepfake scanner used by Tempo Cek Fakta on the Dr Terawan diabetes-drug deepfake and by Fact Crescendo Sri Lanka on the Dhammika Perera deepfake. Documented to flip to "no deepfake detected" on a reduced-resolution, edited Obama clip. Treat the verdict as one weak signal only.
What it does¶
Deepware Scanner accepts a video file or a URL pointing to one and returns a probability score that the video has been manipulated. The web interface is at scanner.deepware.ai; a mobile companion extends the same workflow to a phone. Output is a single number between 0 and 1 with a verdict label ("deepfake detected" / "no deepfake detected"). The tool is open-source-oriented in that underlying models and some tooling are public, with commercial APIs and a low-cost paid tier on top.
The toolkit ships Deepware as the alternative entry in 1A.2 First-Line Triage video alongside InVID-WeVerify (which is the primary because of its source-history pillar coverage). Deepware is the press-button option when the user just wants a directional detector signal in the FLT five-minute window. The video detector class is structurally constrained: regional research found the best commercial video detector in the Deepfake-Eval-2024 anonymised pool reaching only 0.78 accuracy on in-the-wild deepfakes. The Deepware verdict is treated explicitly as one weak signal under Anchor 1, not as a verdict.
When to use it¶
- A short video clip in your queue needs a fast detector signal inside the five-minute first-line window before you decide whether to escalate to InVID-WeVerify for a thirty-minute desk pass.
- You want a press-button option that requires no install or account, accessible on a phone in the field.
- You are running a Tempo or Fact Crescendo SL-style triage workflow where the case has already been documented to use Deepware as the first detector pass.
- A workshop scenario calls for demonstrating the qualitative failure mode (compressed / cropped video flips a deepfake to "no deepfake detected") that the WITNESS / Reuters Institute documented as a baseline limitation of the video detector class.
Independent accuracy¶
Vendor claim vs independent assessment
Vendor claim: Deepware does not publish a sharp headline accuracy figure; mid-tier accuracy in the 80-85% range is on record, characterised by Deepware as subject to "constant adversarial obsolescence."
Independent finding: WITNESS / Reuters Institute's April 2024 methodology piece tested public deepfake detectors on known examples, including a well-known Obama deepfake. Deepware judged a reduced-resolution, edited version of the Obama clip as "No Deepfake Detected" – a qualitative robustness failure, not a benchmark score. Regional research records this as the principal independent observation on Deepware.
No independent SEA-specific benchmark identified as of May 2026. The WITNESS / Reuters Institute test was an example-based audit, not a regional benchmark; no Asian-face or SEA-language breakdown is published.
Limitations¶
Limitations
- Mid-tier accuracy in the 80-85% range; detection capabilities face constant adversarial obsolescence.
- Documented to flip to "No Deepfake Detected" on a reduced-resolution, edited Obama clip.
- SEA-platform-compressed-video benchmark missing ; WhatsApp, Facebook, and TikTok all transcode video on upload, so Deepware verdicts on platform-circulated video carry the asymmetric robustness risk.
- No published per-language or Asian-face breakdown.
Privacy and threat model¶
The web scanner uploads the video file to Deepware's cloud infrastructure for analysis. Deepware is a US-jurisdiction vendor; treat any upload as a transmission to a US server with the vendor's retention and disclosure policy applying. For routine triage on already-public video (a clip already widely circulated on TikTok or Facebook), the upload is low-risk because the content is already public.
For source-identifying video, the upload itself is the risk regardless of the verdict; in surveillance-environment contexts (Sri Lanka, Lao), classify the source file before any upload. If the file is source-identifying, route to InVID-WeVerify for source-history work first (reverse-image, archived versions, metadata) and skip the detector tab.
Source-protection override (S1 — source-identifying upload risk)
Uploading source-identifying video to Deepware's web scanner transmits the file to a US-jurisdiction vendor under that vendor's retention policy. Mitigation steps:
- Classify the video as public, sensitive, or source- identifying before any upload.
- For source-identifying material, do not invoke Deepware. Use InVID-WeVerify's non-uploading keyframe and reverse-image modules for the source-history layer.
- If a Deepware verdict is operationally necessary, strip identifying context (crop the frame, blur backgrounds, remove audio) before uploading the redacted version.
- If you uploaded by mistake, request vendor deletion via the Deepware support contact and disclose to the source.
Source-protection override (S9 — cross-border data transfer)
Deepware operates from US jurisdiction; uploads cross into the US data-protection regime. Mitigation steps:
- Check organisational policy on US-vendor uploads before any Deepware call.
- Record the vendor's retention statement at the time of upload with the case file; vendor terms can change.
- Prefer offline or EU-jurisdiction alternatives (InVID's deepfake tab via CERTH for redacted material with documented 30-day retention) when the source is in a surveillance-risk country.
Country and platform applicability¶
- Indonesia: documented use by Tempo Cek Fakta on the Dr Terawan diabetes-drug deepfake.
- Laos: language-agnostic; not benchmarked on Lao-region content.
- Malaysia: no documented use recorded.
- Philippines: no documented Rappler / VERA Files deployment recorded; the Philippines pipeline routes through Sensity at institutional tier and Hive at first-line.
- Sri Lanka: documented use by Fact Crescendo Sri Lanka on the Dhammika Perera deepfake.
- Thailand: no documented use recorded.
Platform applicability: works on any video file regardless of source platform. The WITNESS / Reuters finding above means the verdict should be treated as weakly informative for platform-compressed content (TikTok, Facebook, Facebook Groups, WhatsApp).
How to access¶
The web scanner is at scanner.deepware.ai. No account is required for free-tier scanning. A mobile companion extends the same workflow to a phone. Commercial APIs are documented at the Deepware developer portal; the paid tier is approximately $8/month .
Cost (current as of 2026-05)¶
Free for the public web scanner; commercial APIs available on a paid basis; the paid tier is roughly $8/month . Verify directly before any institutional integration as vendor pricing may have shifted.
Quickstart¶
- Open scanner.deepware.ai in your browser (works on phone or desktop).
- Upload the video file or paste a URL pointing to it.
- Wait for the analysis; Deepware returns a single probability score with a verdict label.
- Read the score; if the score is high but the file is platform-compressed (TikTok, Facebook, WhatsApp), treat the verdict as one weak signal and escalate to InVID-WeVerify for source-history work.
- If the score is low and the file is platform-compressed, do NOT treat the verdict as exoneration. The WITNESS / Reuters audit shows compression flips deepfake verdicts to "no deepfake detected" on documented examples.
- Pair the result with a non-detector signal (reverse-image hit, archived earlier instance, provenance manifest) before any conclusion.
In the toolkit's workflow¶
Cautious-detector pillar tool per Architectural Anchor 1. Sits in 1A.2 First-Line Triage video as the alternative to InVID-WeVerify. Deepware is the press-button detector option; InVID-WeVerify is the multi-pillar default.
Standard combinations:
- With InVID-WeVerify at 1A.2 / 1B.1 as the source-history primary; Deepware gives the detector signal, InVID-WeVerify gives the source-history signal. Per Anchor 2 these combine as detector + non-detector: two signal classes.
- With Hive at 1A.1 cross-cell when the same case has both a still image and a video; the two together give parallel detector signals counted as one signal class under Anchor 3.
- With Sensity at 1C.2 for institutional escalation when the case warrants enterprise-grade detection.
This tool's verdict is one weak signal class per Anchor 1; never publish a binary claim from it alone. Counted as one detector signal class under Anchor 3 when paired with another detector.
Decision-tree references: T2 video triage – Deepware is the press-button branch at T2.10 (video detector run); the WITNESS / Reuters qualitative failure is documented in the T2 stop-condition annotations on the same node.
Conflict resolution behaviour¶
When Deepware's verdict disagrees with another detector (InVID deepfake tab, Hive on a still frame, Sensity at institutional tier), Deepware carries lower evidence weight on platform-compressed video because the WITNESS / Reuters Institute documented qualitative failure on a known example. On uncompressed video, Deepware carries roughly equal weight to other detectors at First-Line Triage tier; the toolkit's editorial position is that no FLT video detector verdict alone is publishable. When Deepware 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.
Override notes¶
Override notes
-
Weak-signal framing (e4-weak-signal-framing-binding): the workflow section above invokes the detector-as-weak-signal sentence per P3 sub-routine; never publish a binary claim from Deepware alone. The Limitations admonition carries the WITNESS / Reuters qualitative failure verbatim.
-
Vendor wrapping pair (c1-pair-vendor80-vs-regional research-qualitative-fail): Deepware's 80-85% range vendor figure is paired against the WITNESS / Reuters qualitative failure on the reduced-resolution Obama clip. Rendered in full in the Independent Accuracy admonition above.
-
Tempo and Fact Crescendo SL deployment (b4-score-2-Tempo-FactCrescendo-SL): regional documented use across Indonesia and Sri Lanka noted in the Country and platform applicability section; the deployment evidence raises the tool's documented-use score but does not override the e4 weak-signal-framing requirement.
Sources¶
- Deepware. Deepware Scanner — deepfake video detection. Deepware, 2024. deepware.ai.
- Deepware. Deepware Scanner — online video scanning tool. scanner.deepware.ai.
- DW Innovation / EU DisinfoLab. Deepfake Detectors Put to the Test (methodology review cited in card). DW Innovation, November 2025. innovation.dw.com.