Error rates

The measured error rates of the version you are using: validation-set results, per reference set and per camera model, never one number, with how each was measured and what it does not claim. Nothing is sent from this page. We would rather publish a dull number we measured than a bright one we hoped for.

The tables are published in English, as the records are, and name the reference data they describe.

Error-rate pages of earlier releases

PixVeritas — published error rates (ruleset R0, v0.7)

Rates are stated per reference set and per camera model, never as one number, because a single figure would understate the risk on models whose units share a common pattern and overstate it on strong sensors. Two kinds of false match are kept apart: cross-model (an image from a different camera model matching an enrolled unit) and same-model (another unit of the same model matching). A match is stated at Level B (consistent with the camera model). Level C (source-device consistency) is withheld from every model for now (ADR 0038, below): it needs a unit that passed the two-part quality gate, enrolment frames that agree with each other at the median as strongly as the class threshold asks (since engine 0.7.13), a model with a measured profile and class threshold, and a reference set that shows its operating point holding where Level C would be claimed. No level above C (individual-sensor attribution) is ever displayed.

The interval is a cluster bootstrap over units — 1,000 resamples, drawing whole cameras with replacement — because the trials are not independent: comparisons re-use references and images, and every frame from one camera shares that camera's sensor. An exact binomial interval is computed beside it and used as a floor on width, never as the answer. This replaces the Wilson intervals published until 2026-09-20, which were about 1.7× too narrow: on one measured set the per-unit false-alarm rate ran from 0 % to 8.9 %, and pooling the files pretended thousands of independent trials where there were 35 correlated groups. Rates still marked pending re-measure carry the older pooled interval until their run is repeated.

Every number below is computed from the public reference sets named at the end of this page by the project's measurement tooling, from per-comparison records kept in the project's repository at the paths this page names. That repository is not public yet, so a reader cannot rerun them today; the lists of the held-out (locked) devices stay private by design, so that no one can fit to them. The "engine path" columns come from the shipped engine itself (enrolment and match exactly as the browser runs them); the "harness" columns from the Phase 0 measurement run the ruleset was consolidated from.

Operating points per reference set (validation split, 1024 × 1024 analysis window)

Reference setEnrolment protocolDecision thresholdLevel C class thresholdUnitsFNMR (engine path)FNMR (harness)Cross-model FMR (engine)Cross-model FMR (harness)Same-model FMR (engine)Same-model FMR (harness)
VISION (2017)flat-2544.7100092.22 % [0.00–5.59] (4/180)
9 units, 8 models
1.67 % [0.00–4.79] (3/180)
9 units, 8 models
0.00 % [0.00–0.27] (0/1360)
9 units, 8 models
0.21 % [0.00–0.62] (3/1400)
9 units, 8 models
0.00 % [0.00–8.81] (0/40)
2 units, 1 model, 1 same-model pair
indicative (n < 100)
0.00 % [0.00–8.81] (0/40)
2 units, 1 model, 1 same-model pair
indicative (n < 100)
FloreView (2023)flat-2550.0not measured (Level B only)114.55 % [0.45–10.00] (10/220)
11 units, 10 models
4.55 % [0.91–9.55] (10/220)
11 units, 10 models
0.00 % [0.00–0.17] (0/2129)
11 units, 10 models
0.05 % [0.00–0.26] (1/2160)
11 units, 10 models
0.00 % [0.00–9.03] (0/39)
2 units, 1 model, 1 same-model pair
indicative (n < 100)
0.00 % [0.00–8.81] (0/40)
2 units, 1 model, 1 same-model pair
indicative (n < 100)

VISION (2017)'s decision threshold (44.7) is held, not refitted. Refitted with each camera's photographs spread evenly over its files (2026-10-01), the usual rule would set it lower, where the validation cameras show about twice the target false-match rate; at the held value both the training and the validation cameras stay under it. The choice was made after looking at the validation cameras, so for this threshold they are no longer an untouched test: the sealed set is, at v1.0 (reports/phase0/spread-remeasure.md).

The same rates inside the standard operating conditions. The table above counts every test picture; the conditions (below) are what its rates are said to describe. Each test picture was checked as the Workbench checks it (vault-eval --soc), and the rates below are over the pictures inside all four. A picture can miss more than one condition.

Reference setTest pictures inside / checkedMissed (a) quality / (b) size / (c) signal / (d) classFNMRCross-model FMRSame-model FMR
VISION (2017)173 / 1800 / 1 / 6 / 01.73 % [0.00–4.98] (3/173)
9 units, 8 models
0.00 % [0.00–0.28] (0/1307)
9 units, 8 models
0.00 % [0.00–9.49] (0/37)
2 units, 1 model, 1 same-model pair
indicative (n < 100)
FloreView (2023)170 / 2200 / 7 / 3 / 405.88 % [0.00–13.07] (10/170)
9 units, 8 models
0.00 % [0.00–0.23] (0/1632)
9 units, 8 models
0.00 % [0.00–9.74] (0/36)
2 units, 1 model, 1 same-model pair
indicative (n < 100)

With every natural photograph of each validation camera (2026-10-17). The table above takes 20 photographs a camera, spread evenly over its files; every natural photograph gives five to ten times the trials, and is what the sealed sets will be read with (docs/analysis-plan.md).

Reference setProfileTest pictures (inside the conditions)FNMRFNMR inside the conditionsCross-model FMRSame-model FMR
VISION (2017)1,024 × 1,0241,937 (1,859)1.45 % [0.24–3.22] (28/1937)
9 units, 8 models
1.40 % [0.16–3.23] (26/1859)
9 units, 8 models
0.00 % [0.00–0.03] (0/14688)
9 units, 8 models
0.00 % [0.00–0.86] (0/428)
2 units, 1 model, 1 same-model pair
FloreView (2023)1,024 × 1,0241,176 (904)10.63 % [4.62–17.97] (125/1176)
11 units, 10 models
12.06 % [3.67–21.64] (109/904)
9 units, 8 models
0.00 % [0.00–0.03] (0/11365)
11 units, 10 models
0.00 % [0.00–1.76] (0/208)
2 units, 1 model, 1 same-model pair
VISION (2017)2,048 × 2,0481,347 (1,277)0.74 % [0.00–2.09] (10/1347)
6 units, 5 models
0.63 % [0.00–1.95] (8/1277)
6 units, 5 models
0.00 % [0.00–0.06] (0/6307)
6 units, 5 models
0.00 % [0.00–0.86] (0/428)
2 units, 1 model, 1 same-model pair
FloreView (2023)2,048 × 2,0481,072 (824)8.12 % [0.75–16.39] (87/1072)
10 units, 9 models
10.19 % [1.89–20.36] (84/824)
8 units, 7 models
0.00 % [0.00–0.04] (0/9265)
10 units, 9 models
0.00 % [0.00–1.76] (0/208)
2 units, 1 model, 1 same-model pair
VISION (2017)whole frame1,937 (1,859)0.52 % [0.00–1.47] (10/1937)
9 units, 8 models
0.43 % [0.00–1.34] (8/1859)
9 units, 8 models
0.00 % [0.00–0.03] (0/14268)
9 units, 8 models
0.00 % [0.00–0.86] (0/428)
2 units, 1 model, 1 same-model pair
FloreView (2023)whole frame1,176 (904)7.74 % [2.14–14.80] (91/1176)
11 units, 10 models
9.62 % [1.94–18.20] (87/904)
9 units, 8 models
0.01 % [0.00–0.05] (1/11265)
11 units, 10 models
0.00 % [0.00–1.76] (0/208)
2 units, 1 model, 1 same-model pair

FloreView (2023), 1,024 × 1,024: the FNMR is above the 5 % target. Its misses gather on a few phones (Samsung GalaxyA40 34 %, Huawei P30Lite 34 %, Huawei P9Lite 17 % of their photographs), which 20 photographs a camera under-sampled. A miss on a camera original reads NO MATCH, so on such a phone a NO MATCH says less than the table above suggests.

FloreView (2023), 2,048 × 2,048: the FNMR is above the 5 % target. Its misses gather on a few phones (Samsung GalaxyA40 32 %, Huawei P30Lite 31 %, Huawei P9Lite 15 % of their photographs), which 20 photographs a camera under-sampled. A miss on a camera original reads NO MATCH, so on such a phone a NO MATCH says less than the table above suggests.

FloreView (2023), whole frame: the FNMR is above the 5 % target. Its misses gather on a few phones (Huawei P30Lite 30 %, Samsung GalaxyA40 28 %, Huawei P9Lite 17 % of their photographs), which 20 photographs a camera under-sampled. A miss on a camera original reads NO MATCH, so on such a phone a NO MATCH says less than the table above suggests.

The 1,024 × 1,024 and 2,048 × 2,048 rows compare each test picture as the product does (since 2026-10-22): the whole picture, with what its own file records and the profile's frame. Another camera's picture narrower than the profile's frame goes to the searched route, as it does in the product, so each false-match rate above counts every comparison on the route the product takes. Until then these runs cut each test picture to the profile's window and compared every one at the camera's own scale (found 2026-10-19). Of the impostor comparisons the product makes elsewhere, none of them matched.

Reference setProfilePhotographs a cameraImpostor comparisonsOf them off the camera's own scaleMatched there
VISION (2017)1,024 × 1,024201,400480 (34 %)0
FloreView (2023)1,024 × 1,024202,168913 (42 %)0
VISION (2017)1,024 × 1,024every15,1164,908 (32 %)0
FloreView (2023)1,024 × 1,024every11,5734,829 (42 %)0
VISION (2017)2,048 × 2,048every6,7352,045 (30 %)0
FloreView (2023)2,048 × 2,048every9,4733,791 (40 %)0

The likelihood ratio beside a MATCH (ruleset 22)

Beside a MATCH at the camera's own scale the page gives the score's likelihood ratio (LR): how much more often a score this high comes from the enrolled camera's own photographs than from other cameras', for H1 "the enrolled camera took the picture" against H2 "another camera did". Each profile size has its own calibration, fitted on the validation comparisons with every natural photograph that the engine makes at the camera's own scale (another camera's narrower picture goes to the searched route, which gives no LR, and is not counted: re-fitted so on 2026-10-19): at every score the lower of the two reference sets' curves, so it never says more than either supports, held within a third of the smaller set's impostor comparisons (at that cap the page says "at least", its figure rounded down). The words are the evaluative scale's (weak from 2, moderate from 10, moderately strong from 100, strong from 1,000). A NO MATCH, an INCONCLUSIVE, a picture compared at another scale and pixels the browser decoded with its own HEIC or AVIF decoder carry no LR. Weak support for H1 is not support for H2, and the LR is not the probability that the image is authentic.

ProfileLR at a score of 44.7449.96100200CapLR = 1 at a score of
1,024 × 1,02413201511,1522,24825.3
2,048 × 2,0481.21.610671,56341.9
whole frame0.50.63.8232,21559.9

On each reference set's own comparisons, the LR's cost (Cllr: 0 is perfect, 1 says nothing; beside it the best any calibration of the same scores that keeps their order reaches), its misleading evidence, the LRs beside the MATCHes, and the share of genuine and of impostor comparisons whose LR reaches each power of ten:

Reference setProfileCllr (best)Genuine with LR < 1Impostors with LR > 1MATCHes: genuine, at the capMATCHes: impostors (LR)LR ≥ 1LR ≥ 10LR ≥ 100LR ≥ 1,000
VISION (2017)1,024 × 1,0240.034 (0.027)12 of 1,93715 of 10,2081,909, 1,731099.4 % / 0.15 %98.7 % / 0.00 %97.2 % / 0.00 %92.8 % / 0.00 %
FloreView (2023)1,024 × 1,0240.308 (0.190)93 of 1,1689 of 6,7441,047, 802092.0 % / 0.13 %90.2 % / 0.01 %85.4 % / 0.00 %74.5 % / 0.00 %
VISION (2017)2,048 × 2,0480.040 (0.031)10 of 1,3470 of 4,6901,337, 1,254099.3 % / 0.00 %99.1 % / 0.00 %98.1 % / 0.00 %94.1 % / 0.00 %
FloreView (2023)2,048 × 2,0480.257 (0.188)84 of 1,0651 of 5,682981, 810092.1 % / 0.02 %90.6 % / 0.00 %88.6 % / 0.00 %79.2 % / 0.00 %
VISION (2017)whole frame0.028 (0.022)10 of 1,9370 of 8,9481,927, 1,829099.5 % / 0.00 %99.4 % / 0.00 %99.1 % / 0.00 %96.1 % / 0.00 %
FloreView (2023)whole frame0.267 (0.173)93 of 1,1680 of 6,6441,080, 984092.0 % / 0.00 %90.9 % / 0.00 %89.6 % / 0.00 %86.7 % / 0.00 %

Each Tippett cell is genuine / impostor.

The decision itself has an LR too, apart from the score's: how much more often a MATCH comes from the enrolled camera than from another, (1 − FNMR) / FMR at the operating point of the tables above (20 photographs a camera). No false match was counted there, so the FMR is taken at its exact upper bound and the LR is a lower bound:

Reference setProfileDecision thresholdFNMRFalse matchesFMR at mostDecision LR at least
VISION (2017)1,024 × 1,02444.742.22 %00.40 %244
VISION (2017)2,048 × 2,04844.741.67 %00.87 %112
VISION (2017)whole frame44.741.11 %00.46 %215
FloreView (2023)1,024 × 1,02449.964.57 %00.29 %325
FloreView (2023)2,048 × 2,04849.965.00 %00.69 %137
FloreView (2023)whole frame49.963.64 %00.59 %162

Beside a MATCH the LR can be below 1, where the curve crosses 1 above the decision threshold: FloreView (2023), whole frame: 5 genuine MATCHes. From 0.5 to 2 the page says the score supports neither proposition; below 0.5, that it supports H2.

Where the operating point can be demonstrated

The Level C operating point is required to hold in stratum 1 (another unit of the same model) and stratum 4 (computational-photography pipelines), not merely pooled across everything. Reporting one pooled false-match rate hides exactly the failure mode that matters on recent phones.

Reference set1 same-model2 cross-model3 classic pipelines4 computational photography
VISION (2017)0.00 % (0/40, 2 units)
Indicative
0.00 % (0/1360, 9 units)
Holds
0.00 % (0/1400, 9 units)
Holds
— NotPopulated
FloreView (2023)0.00 % (0/39, 2 units)
Indicative
0.00 % (0/2129, 11 units)
Holds
— NotPopulated0.00 % (0/2168, 11 units)
Holds

Holds means the whole interval sits at or below the 0.5 % target; ConsistentWith means the rate does but the interval does not; AboveTarget means the rate is above the target but the interval reaches below it; Fails means the whole interval is above the target; Indicative means fewer than 100 comparisons or too few units to resample; NotPopulated means the set contributes no comparison of that kind.

VISION (2017). stratum 1 is indicative: 40 comparisons from 2 device(s), under the 100-trial floor; stratum 4 is not populated: VISION is a classic-pipeline source, so it holds no computational-photography comparison; that stratum is measured on the computational-photography sets.

FloreView (2023). stratum 1 is indicative: 39 comparisons from 2 device(s), under the 100-trial floor.

What a rescaled copy costs

A picture that has been through a messaging app or a social platform arrives rescaled. These are the false-non-match rates when the whole frame is rescaled by the factor shown and the geometry search is left to recover it — the rate that applies to a redistributed copy, rather than to a file straight off the camera.

Reference setRescaleMissedDeclined to judge
VISION (2017)1.00×2.22 % [0.00–5.59] (4/180)0
VISION (2017)0.75×2.78 % [0.00–6.67] (5/180)0
VISION (2017)0.50×8.33 % [2.22–17.22] (15/180)0
VISION (2017)0.25×60.00 % [45.00–72.78] (108/180)0
FloreView (2023)1.00×4.11 % [0.45–9.55] (9/219)1
FloreView (2023)0.75×6.39 % [2.28–11.42] (14/219)1
FloreView (2023)0.50×13.76 % [5.53–23.29] (30/218)2

Since ruleset 19 a miss here can also be a match the engine withholds (ADR 0042): its score cleared the threshold, but most of it came from the finest detail, as a picture enlarged after it was taken scores against another camera's profile, so it is reported as inconclusive and counted as a miss (VISION (2017): 2 of the 108 misses at 0.25×; FloreView (2023): 3 of the 30 misses at 0.50×). Here they are the table's smallest comparisons, but the size does not bound them: on the reference sets' training devices the engine withholds such matches at up to about 2 megapixels compared, mostly on two cameras.

The rate tracks how much picture is left to compare. Through a 1024 x 1024 analysis window a half-size copy leaves a 512-pixel square to correlate and a quarter-size copy leaves 256, which is why the last row is so much worse than the others rather than a little worse. A profile enrolled from the whole frame keeps far more at the same rescale, so these figures are the window's rates, not the method's ceiling.

"Declined to judge" counts comparisons the engine refused because the geometry fell outside the range it has been validated over. Those are excluded from the rate rather than counted as misses: a picture the engine would not judge is not a picture it got wrong, and folding the two together is how a detector comes to look worse — or better — than it is.

Another camera's photograph of the same size: the profile's size (ADR 0016)

A picture at the enrolled camera's own scale that holds the profile's whole frame is compared once, at the camera's own threshold, which was fitted on profiles of the 1,024² window. Until engine 0.7.9 the score was the strongest correlation anywhere in the comparison, and against a whole-frame profile, the Vault's default, another camera's photograph of the same size then cleared that threshold more often than 0.5 % (the counts are in docs/vis/ruleset-14.md). Since engine 0.7.9 the score is taken at the alignment the geometry implies (ADR 0016), and on it the threshold fitted for each profile size is 24.18 for the 2,048² window, or the whole frame of a camera of up to 4.6 MP and 22.00 for the whole frame of a larger camera (mid: VISION 19.77, FloreView 24.18; large: VISION 19.51, FloreView 21.99): below every camera's threshold, so the camera's own threshold applies at every size. Measured through the product (Stage 6 profiles, Stage 7 deciding every comparison), with comparisons against a profile the shared-pattern rule withholds left out. A photograph stored turned against the profile's frame is compared turned on the searched route, which the size thresholds do not govern, and is left out too (its stored size tells it).

ProfileReference set, splitAt the camera's thresholdAt the size thresholdRight camera found
2,048² window, or the whole frame of a camera of up to 4.6 MPVISION, training0.00 % [0.00–0.42] (0/869, 8 devices)0.00 % [0.00–0.42] (0/869, 8 devices)68.64 % [44.44–88.61] (116/169, 9 devices)
2,048² window, or the whole frame of a camera of up to 4.6 MPVISION, validation (held out)0.00 % [0.00–0.87] (0/420, 5 devices)0.00 % [0.00–0.87] (0/420, 5 devices)98.33 % [94.11–100.00] (118/120, 6 devices)
2,048² window, or the whole frame of a camera of up to 4.6 MPFloreView, training0.00 % [0.00–0.13] (0/2890, 18 devices)0.00 % [0.00–0.13] (0/2890, 18 devices)83.68 % [68.42–95.26] (159/190, 19 devices)
2,048² window, or the whole frame of a camera of up to 4.6 MPFloreView, validation (held out)0.00 % [0.00–0.69] (0/530, 10 devices)0.00 % [0.00–0.69] (0/530, 10 devices)95.00 % [88.00–100.00] (95/100, 10 devices)
the whole frame of a larger cameraVISION, training0.00 % [0.00–0.27] (0/1374, 12 devices)0.00 % [0.00–0.27] (0/1374, 12 devices)78.31 % [60.64–94.62] (195/249, 13 devices)
the whole frame of a larger cameraVISION, validation (held out)0.12 % [0.00–0.69] (1/800, 8 devices)0.12 % [0.00–0.69] (1/800, 8 devices)98.89 % [96.04–100.00] (178/180, 9 devices)
the whole frame of a larger cameraFloreView, training0.00 % [0.00–0.15] (0/2470, 19 devices)0.00 % [0.00–0.15] (0/2470, 19 devices)81.50 % [64.50–95.00] (163/200, 20 devices)
the whole frame of a larger cameraFloreView, validation (held out)0.00 % [0.00–0.65] (0/570, 10 devices)0.00 % [0.00–0.65] (0/570, 10 devices)95.45 % [89.09–100.00] (105/110, 11 devices)

Redistributed copies: the search's own threshold

To test a rescaled picture the engine tries the sizes the picture's own dimensions imply, in each turn its shape allows, and keeps the best score. A picture that is not a whole-frame resize of the enrolled camera's picture is handled by the crop route below or not tested (ADR 0014, ADR 0015). The best of several comparisons runs higher than a single one even for the wrong camera, so this route has a threshold of its own: the engine applies the higher of the ordinary threshold (VISION 44.74, FloreView 49.96) and 43.63, the lowest score at which the training devices' wrong-camera comparisons on this route stay at or below 0.5 %, taken over both kinds of another camera's picture that reach it (VISION: another camera's messenger copies 39.74; VISION: another camera's photographs at a smaller size 43.63; FloreView: another camera's WhatsApp-like copies 20.55; FloreView: another camera's photographs at a smaller size 28.34).

Re-fitted in ruleset 14 (2026-10-02), on both reference sets and on this route alone. Ruleset 13's 50.66 was fitted on VISION over rows that included another camera's photographs of the same size; those take the native route (next section), where this threshold does not apply, and their tail set the fit. Since ruleset 14 every comparison is counted on the route the engine sends it down, and comparisons against a profile the shared-pattern rule withholds are left out (no MATCH can come of them). The fit is the measuring tool's search, whose VISION rows predate the score the engine takes at the alignment and whose FloreView rows predate the middle it reads of a profile above 24 MP: they hold 16 wrong-camera comparisons at or over the threshold applied. The rates below are the engine's: the same pictures against the same profiles, scored again through it as the product runs it (2026-10-16), each with its own file facts, FloreView's WhatsApp-like copies made as its runs made them (long side 1,600 pixels, saved once at quality 92). On the same comparisons the engine reports 1 false match, and withholds 1 more whose score cleared the threshold mostly in the finest detail (ADR 0042).

Wrong-camera picturesSplitThresholdFalse matchesWithheld (finest detail)
VISION: another camera's messenger copiestraining49.960.00 % [0.00–0.06] (0/6373, 15 devices)0
VISION: another camera's messenger copiesvalidation (held out)49.960.00 % [0.00–0.15] (0/2400, 9 devices)0
VISION: another camera's photographs at a smaller sizetraining44.74, 49.960.00 % [0.00–0.30] (0/1211, 13 devices)0
VISION: another camera's photographs at a smaller sizevalidation (held out)44.740.00 % [0.00–1.15] (0/320, 7 devices)0
FloreView: another camera's WhatsApp-like copiestraining49.960.00 % [0.00–0.11] (0/3230, 17 devices)0
FloreView: another camera's WhatsApp-like copiesvalidation (held out)49.960.00 % [0.00–0.50] (0/740, 11 devices)0
FloreView: another camera's photographs at a smaller sizetraining49.960.08 % [0.00–0.47] (1/1190, 17 devices)1
FloreView: another camera's photographs at a smaller sizevalidation (held out)49.960.00 % [0.00–1.08] (0/340, 9 devices)0

Only comparisons the engine searched are counted: a picture of another shape is not on this route. The threshold was not re-fitted on the held-out devices, because a threshold chosen on the data that tests it is not tested. Intervals are clustered by enrolled camera, which is why they are wide on nine devices.

How many copies of the right camera it finds, through the engine as the Vault tests them: VISION's WhatsApp and Facebook copies (made by those platforms in 2016–17 from photographs of its 2014–16 phones), each with its own file facts, against the camera's profile enrolled from flat fields as the Vault asks. Copies the engine does not test (below a quarter of the enrolled size, or of another shape) are left out.

ProfileSplitWhatsAppFacebook (high)Facebook (low)
whole frametraining80.36 % [71.79–88.93] (225/280, 14 devices)60.00 % [43.67–76.33] (180/300, 15 devices)14.67 % [2.67–30.33] (44/300, 15 devices)
whole framevalidation96.88 % [92.86–100.00] (155/160, 8 devices)80.56 % [61.67–94.44] (145/180, 9 devices)8.89 % [3.33–16.67] (16/180, 9 devices)
1024 x 1024 windowtraining28.08 % [12.31–45.77] (73/260, 13 devices)17.14 % [5.00–29.64] (48/280, 14 devices)0.71 % [0.00–2.56] (2/280, 14 devices)
1024 x 1024 windowvalidation41.25 % [17.50–65.00] (66/160, 8 devices)19.44 % [3.33–42.22] (35/180, 9 devices)0.00 % [0.00–2.03] (0/180, 9 devices)

A profile of the whole frame finds most WhatsApp and high-quality Facebook copies; one enrolled through the 1024 x 1024 window finds about a third of the WhatsApp copies and a fifth of the Facebook ones, because a rescaled copy leaves only a small square of that window to correlate. Facebook's low-quality copies are rarely found either way. A copy that is not found is reported as not tested, never as a different camera: a re-encode removes the signature whether or not the camera matches.

Pictures cut from the camera's frame (ADR 0015)

A picture that no resize of the enrolled camera's whole picture explains is tried once more, as a piece cut from that picture at the camera's own scale, in each turn it fits: an in-camera square or wide mode, or a crop saved without resizing. That one geometry has its own threshold, because a square can be tried four ways and the best of four runs higher for the wrong camera: 58.72, fitted on the training devices of VISION and FloreView at a false-match rate of at most 0.5 % (the engine applies VISION 58.72, FloreView 58.72). A piece that is not found is reported as not tested, never as a different camera, because a crop that was also resized misses the same way. A file with a messenger or social-media signature is not tried this way at all: a platform resizes, so the camera's own scale is not a geometry such a copy can have.

The mirror check (ADR 0017). A piece's place in the frame is unknown, so this route takes the strongest correlation anywhere in the picture, and a picture made mostly of fine texture (stone walls, paving) can score high there against any profile at all: on FloreView's training devices one such photograph cleared the threshold against 9 of the 16 other cameras' profiles it was tried against, and with the check against 0. So a score that clears the threshold must also reach 1.46 times what the same picture scores against the profile flipped into its mirror image, which no camera produces; otherwise the result is inconclusive. The factor is fitted like every threshold here, on the training devices of both sets at a false-match rate of at most 0.5 %, for the check on its own.

Wrong-camera pictures it can meetSplitFalse matchesWith the mirror check
VISION: square crops of another camera's photographstraining0.05 % [0.00–0.25] (1/2200, 14 devices)0.00 % [0.00–0.17] (0/2200, 14 devices)
VISION: square crops of another camera's photographsvalidation (held out)0.00 % [0.00–0.44] (0/840, 8 devices)0.00 % [0.00–0.44] (0/840, 8 devices)
VISION: another camera's photographs of another shapetraining0.00 % [0.00–0.59] (0/620, 10 devices)0.00 % [0.00–0.59] (0/620, 10 devices)
VISION: another camera's photographs of another shapevalidation (held out)0.00 % [0.00–1.53] (0/240, 6 devices)0.00 % [0.00–1.53] (0/240, 6 devices)
FloreView: square crops of another camera's photographstraining0.44 % [0.22–0.79] (11/2480, 19 devices)0.04 % [0.00–0.22] (1/2480, 19 devices)
FloreView: square crops of another camera's photographsvalidation (held out)0.34 % [0.00–1.22] (2/590, 10 devices)0.34 % [0.00–1.22] (2/590, 10 devices)
FloreView: another camera's photographs of another shapetraining0.00 % [0.00–1.66] (0/220, 19 devices)0.00 % [0.00–1.66] (0/220, 19 devices)
FloreView: another camera's photographs of another shapevalidation (held out)0.00 % [0.00–2.15] (0/170, 9 devices)0.00 % [0.00–2.15] (0/170, 9 devices)

What the check cannot address is a pattern two particular phones share, which the mirror does not carry: the 2 held-out false matches it leaves are of that kind, one pair of phones, and stay open (reports/phase0/crop-decoys.md, section 5).

Kept out by that rule, VISION: another camera's messenger copies of another shape would have read 0.26 % [0.00–0.61] (5/1920, 9 devices) held out at the same threshold, 4 of them against one enrolled camera, and the route found 0 of the 40 right-camera copies it tried there.

How many pieces of the right camera's pictures it finds (VISION 58.72, FloreView 58.72), with the profile enrolled from flat fields; the mirror check withheld 1 of the 2,272 it found:

PieceTrainingValidation
VISION: square, centred94.00 % [89.33–98.00] (282/300, 15 devices)98.89 % [96.04–100.00] (178/180, 9 devices)
VISION: square, at the edge92.33 % [87.67–96.33] (277/300, 15 devices)98.89 % [96.04–100.00] (178/180, 9 devices)
VISION: 4:5 portrait94.00 % [89.67–97.67] (282/300, 15 devices)98.89 % [96.04–100.00] (178/180, 9 devices)
VISION: square, saved at q9092.00 % [87.33–96.67] (276/300, 15 devices)98.89 % [96.04–100.00] (178/180, 9 devices)
VISION: square, saved at q75not measured98.89 % [96.04–100.00] (178/180, 9 devices)
FloreView: square, centred80.50 % [63.50–94.00] (161/200, 20 devices)94.55 % [85.45–100.00] (104/110, 11 devices)

The pieces are cut from VISION's and FloreView's photographs by the measurement, not by a person, and only the q90 and q75 rows were saved again as JPEG before testing. The held-out devices were examined before the threshold was fitted, so they do not test it untouched; the locked set does, at v1.0, and so will the founder's own photographs.

Standard operating conditions

The rates above apply only to a picture that meets all four of these, which are recorded in the signed ruleset:

A picture that misses any of them is reported outside these conditions: the finding still stands, but no rate on this page describes it. Lossless input (PNG, TIFF) and platform-decoded HEIC/AVIF have no validated conditions at all, because condition (a) is defined for JPEG. A file carrying a messenger or social-media re-encode signature is tested as a copy ("Redistributed copies" above): a MATCH is reported, and a miss reads not tested, never a different camera, because a re-encode removes the signature whether or not the camera matches. The Workbench names the signature before it runs.

Scope, in plain counts

Quality gate (Step 9) floors

Model profiles

One row per camera model measured with at least two units. The ceiling is the highest attribution level the engine may state for an admitted unit of that model: C when the same-model false-match rate is at or below 2 %, otherwise B. Single-unit models are Level B only and are not listed.

ModelReference setUnitsSame-model FMRCross-model FMRShared patternCeilingNote
Apple iPhone 5VISION (2017)20.00 % [0.00–8.81] (0/40)
2 units, 1 model, 1 same-model pair
indicative (n < 100)
0.71 % [0.00–2.56] (2/280)
2 units, 7 models
noBmeasured ceiling C; Level C withheld until strata 1 and 4 hold (ADR 0038)
Google Pixel 3aFloreView (2023)20.00 % [0.00–8.81] (0/40)
2 units, 1 model, 1 same-model pair
indicative (n < 100)
0.00 % [0.00–1.02] (0/360)
2 units, 9 models
noBmeasured ceiling C, with no Level C class threshold in this reference set; Level C withheld until strata 1 and 4 hold (ADR 0038)
Xiaomi Redmi Note 8TFloreView (2023)20.00 % [0.00–8.81] (0/40)
2 units, 1 model, 1 same-model pair
indicative (n < 100)
0.12 % [0.00–0.69] (1/800)
2 units, 20 models
noBmeasured ceiling C, with no Level C class threshold in this reference set; Level C withheld until strata 1 and 4 hold (ADR 0038)

Level C is withheld (ADR 0038). 3 models measured a Level C ceiling, but Level C is offered only when one reference set with a Level C class threshold shows stratum 1, stratum 4 and the FNMR at its operating point all holding (ADR 0043; the strata table above). Today stratum 1 (another unit of the same model) holds in no reference set yet, stratum 4 (computational-photography pipelines) holds only in FloreView (2023), which has no Level C class threshold, and the FNMR at the operating point (its whole interval at or below 5 %) holds in no reference set yet, so every model's ceiling is B. The measured ceiling stays in the ruleset, so a ruleset refresh restores Level C once one set shows all three.

3 models with two or more units; 0 with a Level C ceiling in force (the measured ones are withheld, above), of which 0 can reach Level C. A ceiling is necessary, not sufficient: the reference set must also have measured a Level C class threshold (a score set from its comparisons when the ruleset was consolidated, not the likelihood ratio given beside a MATCH), and the score must clear it. Models measured in more than one reference set keep the profile with more units in the shipped engine; the others remain here for the record.

Spatial integrity (Layer 4)

False alarms are measured on unedited photographs - genuine camera frames re-saved across two encoders, eight qualities, three chroma modes and both scan types, from the cameras of the training and validation splits (the locked split's cameras are kept for the v1.0 evaluation), and the same cameras' photographs as they wrote them. Every one of them is untouched, so any region reported is wrong, and there is no ground truth to argue about. Detection is measured on separate sets of edits built from the same licence-clean source sets; "located" means a reported region overlaps the edited area, because a detector that fires elsewhere in the frame has not found anything. VISION is a set published in 2017; FloreView one published in 2023, with phones up to 2021.

Both encoders are libjpeg-derived, so this measures settings variety honestly and encoder-family variety not at all: Apple's, Adobe's and the phone makers' own JPEG writers are not represented.

DetectorReference setFalse alarms on untouched photographs re-savedAssessedWhat it declines to judge
Noise-field continuityVISION1.96 % [0.98-3.12] (94/4800, 24 units)4.59 % [2.55-6.89] (94/2049)a picture too textured for half its blocks to be measurable (57 % of files)
Noise-field continuityFloreView3.38 % [2.27-4.70] (162/4800, 33 units)9.18 % [6.62-12.17] (162/1764)a picture too textured for half its blocks to be measurable (63 % of files)
Rescaling signatureVISION0.12 % [0.00-0.30] (6/4800, 24 units)0.26 % [0.00-0.61] (6/2347)any picture saved below encoder quality 90 - the trace is erased (51 % of files)
Rescaling signatureFloreView0.71 % [0.23-1.27] (34/4800, 33 units)1.45 % [0.47-2.61] (34/2347)any picture saved below encoder quality 90 - the trace is erased (51 % of files)
Duplicated regionVISION0.10 % [0.00-0.25] (5/4800, 24 units)0.11 % [0.00-0.25] (5/4761)a frame too small to tile, or with too little textured content to compare (1 % of files)
Duplicated regionFloreView0.10 % [0.00-0.27] (5/4800, 33 units)0.10 % [0.00-0.27] (5/4800)a frame too small to tile, or with too little textured content to compare (0 % of files)

The third column is what someone submitting a picture experiences. The fourth is the detector's own error rate among the pictures it was willing to judge at all - and the two can differ by more than a factor of two, because a declined picture is not a quiet pass.

The photographs as the phones wrote them (no re-save), each judged with its own encoder quality, as the Workbench does:

DetectorVISIONFloreView
Noise-field continuity2.01 % [0.72-3.83] (101/5025, 24 units)4.53 % [2.88-6.61] (160/3534, 33 units)
Rescaling signature0.58 % [0.22-0.96] (29/5025, 24 units)2.12 % [0.85-3.74] (75/3534, 33 units)
Duplicated region0.14 % [0.00-0.31] (7/5025, 24 units)0.08 % [0.00-0.25] (3/3534, 33 units)
Compression grid continuity0.04 % [0.00-0.14] (2/5025, 24 units)0.57 % [0.09-1.24] (20/3534, 33 units)

The compression grid continuity row was measured at the 128 px cell before this ruleset's 192 px one (the rows record the engine's status, which cannot be re-cut to another cell); its rate on photographs saved twice is given at this ruleset's cell below.

EditVISION: noise-field continuity locatedFloreView: noise-field continuity located
region smoothed7.7 % (69/900)4.6 % (41/900)
region filled from its surroundings6.7 % (60/900)4.9 % (44/900)
region pasted from another camera2.2 % (20/900)2.7 % (24/900)
region pasted and rescaled3.7 % (33/900)4.2 % (38/900)
region copied from elsewhere in the same picture1.0 % (9/900)1.1 % (10/900)
whole frame rescaled (quality 90 and above)76.0 % (456/600) by the rescaling signature71.8 % (431/600) by the rescaling signature

A copied region is near zero by design, not by failure: a copy taken from the same picture carries the same noise (the duplicated-region detector is its check). The noise-field detector locates about one region smoothed or filled in 14 on VISION and 21 on FloreView, so an unmarked picture is not a cleared picture - that sentence appears on every report carrying this finding.

Copies found by the duplicated-region detector (the validation cameras, held out; a copy is found when the reported area joins it to its source). Since ruleset 17 the detector has two routes: matching blocks, which find a region copied unchanged, and matching keypoints, which also follow a copy turned, rescaled or mirrored first (ADR 0036, reports/phase0/copy-move/).

CopyReference setFound
copied unchangedVISION97.0 % (748/771)
copied unchangedFloreView98.9 % (445/450)
copied after a turn of 3 to 90 degreesVISION73.3 % (286/390)
copied after a rescale of 0.8 to 1.25VISION62.8 % (241/384)
copied as its mirror imageVISION59.2 % (190/321)

Small copies (the validation cameras, held out; reports/phase0/copy-move/small-copies.json). A region copied unchanged is found less often the smaller it is, and one covering less than about half a percent of the frame not at all:

Copied unchanged, coveringFound
0.3 % of the frame0.0 % (0/78)
0.5 % of the frame26.1 % (18/69)
1 % of the frame66.7 % (60/90)
2 % of the frame81.6 % (93/114)

A small region copied after a turn, a rescale or a mirror is found at most 11 % of the time.

On recent phones two of these checks misfire. The noise-field continuity check flags an untouched FloreView photograph about as often as it locates an edit, and the compression grid continuity check marks an area on 6.86 % [4.87-8.85] (147/2142, 33 units) of FloreView photographs saved twice without an edit (VISION: 1.31 %; this ruleset's 192 px cells, z 20, regions of 4 cells). For neither does any setting meet its false-alarm target on FloreView while still finding edits (reports/phase0/FloreView/spatial-integrity.md), so a reading from either on a recent phone's picture is weak evidence. Since ruleset 17 both measure and conclude nothing (ADR 0026): each reports INCONCLUSIVE, "measured, not concluded", on every picture, so neither an area flagged nor a clean reading from them is reported as a finding. The rates above are what they measured.

Objects inpainted into a photograph: not found

No check here finds an object a diffusion model has inpainted into a photograph. Measured on TGIF's validation split (CC BY-SA 4.0; its testing split is sealed and unread), pictures at least 600 px on their shorter side, scored with the spatial checks this version runs, under ruleset 18 (whose spatial settings this version's ruleset keeps). The two checks that conclude, the rescaling signature and the duplicated region search, mark an area on at most 0.5 % of these pictures, and on none of the 306 authentic photographs. Intervals are clustered by the source photograph.

InpaintingA concluding check marks an areaThe area is on the objectMeasured, not concluded: noise field / compression grid mark an area
Stable Diffusion 2, spliced into the photograph0.0 % [0.0-0.2] (0/1836)0.0 % [0.0-0.2] (0/1836)0.7 % / 0.0 %
Photoshop / Firefly generative fill, spliced0.5 % [0.1-1.1] (9/1836)0.5 % [0.1-1.1] (9/1836)1.2 % / 0.0 %
FLUX.1 [schnell], spliced0.0 % [0.0-0.2] (0/1836)0.0 % [0.0-0.2] (0/1836)0.9 % / 0.0 %
SDXL, the whole frame regenerated0.2 % [0.0-0.5] (3/1836)0.1 % [0.0-0.4] (2/1836)1.2 % / 0.2 %
FLUX.1 [schnell], the whole frame regenerated0.0 % [0.0-0.2] (0/1770)0.0 % [0.0-0.2] (0/1770)1.4 % / 0.0 %
Authentic photographs (every mark is a false alarm)0.0 % [0.0-1.2] (0/306)-1.0 % / 0.0 %

A diffusion inpaint carries plausible noise of its own and no rescaling, copy or second JPEG grid, so there is nothing in it for these checks to find. A picture whose whole frame was regenerated is a different case: every pixel went through the generator, which is what a check for synthetic pictures would look for, and no such check is part of this version.

Phones whose own processing makes them match each other

Samsung phones with Exynos processors (the Global variants) write a periodic diagonal pattern into every photograph in the default camera mode. Vázquez-Padín and Pérez-González (ACM IH&MMSec 2026) find six such patterns across models from 2018 to 2025: the Galaxy S9 to S10+, S20, S21 (HDR mode), S22 Ultra, S24 and S25 FE series, Note9 and Note10 series, Z Flip7, and the A25 to A56, as well as the Exynos vivo X60. The US Snapdragon variants carry different patterns of their own. Every unit of a model carries it, so the fingerprints of different phones correlate. On our own engine, before the rule below, Galaxy S10+ photographs matched a Galaxy S10 profile 19 times in 20, and S10 photographs an S10+ profile 17 times in 20, at up to 900× the threshold (FloreView train split, product path).

The engine now tests every profile for such a pattern before it reports a match: when the profile's signature repeats itself off the origin (self-correlation ≥ 0.0241 for a profile of the 1,024² window, ≥ 0.01 for a larger one, with score ≥ 60, read on its centred 1024² window), a score above the threshold is reported INCONCLUSIVE, never MATCH. Measured through Stage 7 on the same split: all 36 of those false matches are now INCONCLUSIVE, and no S10/S10+ cross-comparison is a MATCH. Pooled impostor matches on that split fall from 45 of 9,240 to 8 of 9,240. The cost is stated as plainly: the owner of an affected phone cannot get a MATCH for their own photographs either (36 of 40 genuine S10/S10+ tests are INCONCLUSIVE), because the score cannot single out the device. PRO-mode RAW capture avoids the pattern on the S series; a profile enrolled that way is tested normally.

Samsung also marks, in the file itself, a photograph the camera fused from several exposures (HDR and multi-frame processing). Fusion shifts the hardware signature: on the owners' own photographs, 17% of fused frames missed the threshold against 2% of single exposures. For a fused photograph the engine may still report a match, but it never reports NO MATCH; a miss there says nothing about which camera took it.

The floor depends on the profile's size, because the engine's cleaning of a profile (Step 8) flattens a repeating pattern more on a larger one: the same Galaxy S10+ reads 0.167 on a 1,024² profile and about 0.05 on the whole frame, and lower again from ordinary photographs, whose scenes add power. Measured on profiles built as the Vault builds them, at every window it offers, from flat fields and from ordinary photographs: 24 of the 24 Galaxy S10 and S10+ profiles of 15 to 25 frames are withheld, and 0 of the 608 profiles of other cameras. With fewer frames than the Vault asks for, the rule can miss a profile of 4 ordinary photographs at 1,024²; 4 ordinary photographs at 2,048²; 4 ordinary photographs at the whole frame; 8 ordinary photographs at the whole frame.

Calibration: tools/datasets/shared_pattern_fit.py over reports/phase0/shared-pattern/periodic-references.csv: the periodic test (origin neighbourhood 16 px, axis band 3) on product-path references (vault-enrol, Stage 6) at the Vault's three windows, flat and natural, 25 and 15 frames, VISION and FloreView train+validation. Per size class (fingerprint area up to 1,153,434 pixels, and above) the floor is the geometric mean of the highest clean and the lowest affected reading at 15-25 frames: small 0.0241 (clean 0.0118 D23_Google_Pixel3a, affected 0.0492 D30_Samsung_GalaxyS10+ natural-15-1024), large 0.01 (clean 0.0076 D27_DOOGEE_S96Pro, affected 0.0132 D30_Samsung_GalaxyS10+ natural-15-native). Score floor 60: about twice the null's maximum at the 1,024² window; every affected reading at 15-25 frames is above it (lowest 183). Ruleset 8's single floor (normalised correlation 0.05, score 200, 1,024² flat references) missed whole-frame and natural-frame S10/S10+ profiles. Galaxy S20+/S21+ are in FloreView's locked split and were not measured. Source: arXiv 2510.09509.

Also measured on the authors' own phones (Apache-2.0: Galaxy S10 wide and telephoto, two Galaxy S24 units, A54, A56) together with FloreView's S10 and S10+: without the rule, one S24 matched the other in 20 of 20 tests; with it, 0 of 1,099 cross-phone comparisons is a MATCH (337 INCONCLUSIVE). Every one of the eight profiles shows the pattern at least 2.8× above the threshold, for both the first pattern (S10 family) and the newest (S24, A54, A56). Patterns 2 to 5 (S20, S21, S22 Ultra, Note10+, several A models) are not yet measured; FloreView's S20+ and S21+ sit in its sealed locked split. The test assumes no particular lattice, so it should catch them too.

Dark photographs

Under-exposed photographs carry less of a camera's signature, and published work reports more false camera matches on them. The reference sets hold few: of their training and validation photographs, VISION (2017) holds 30 below a mean brightness of 50 (of 255) and 116 below 70; FloreView (2023) holds 0 below a mean brightness of 50 (of 255) and 13 below 70. Every one below 70, and as many ordinary photographs (100 to 140) from the same cameras, was tested against every camera's whole-frame profile as the Vault tests a picture:

Reference setPhotographsOwn camera matchedOther cameras matched
VISION (2017)116 below 7091.38 % [79.17-98.00] (106/116)0.00 % [0.00-0.14] (0/2668)
VISION (2017)30 below 50 indicative (n < 100)90.00 % [61.54-98.31] (27/30)0.00 % [0.00-0.53] (0/690)
VISION (2017)116 from 100 to 140 (same cameras)96.55 % [88.41-100.00] (112/116)0.00 % [0.00-0.14] (0/2668)
FloreView (2023)13 below 70 indicative (n < 100)61.54 % [22.22-100.00] (8/13)0.00 % [0.00-0.88] (0/416)
FloreView (2023)13 from 100 to 140 (same cameras) indicative (n < 100)46.15 % [13.33-84.62] (6/13)0.00 % [0.00-0.88] (0/416)

No other camera matched a dark photograph. On VISION (2017), which holds the most, a camera matched 91.4 % of its own dark photographs against 96.6 % of its ordinary ones. These sets cannot give a rate for genuinely under-exposed photographs as a class, and none is claimed: they will be measured on photographs taken for the purpose.

Their copies were tested too: each as its WhatsApp and two Facebook copies and as an Instagram-style square, against every camera's profile made from natural photographs as well as from flat fields. Other cameras matched 0 of 30,336 comparisons below a mean brightness of 70. Another set of photographs, with pictures darker than any these sets hold, did show other cameras matching very dark pictures (FODB, below). Since this version the camera check holds every match to what the same picture scores against the profile flipped into its mirror image, which no camera produces, as it already did for a piece cut from the camera's picture: a match the mirror nearly reproduces is inconclusive. On these sets' photographs and copies it withheld none of the right camera's 1,319 matches. On FODB, which is never used to set a rule, it withholds 53 of the 84 other-camera matches on dark pictures, 2 of the 4 on pictures inside all four conditions, and 7 of the right camera's 65 matches on the same pictures; the rest stay open until photographs taken in low light on purpose can be measured.

Bright photographs

In an over-exposed photograph the clipped highlights hold none of a camera's signature. The reference sets hold few bright photographs: VISION (2017) holds 32 above a mean brightness of 180; FloreView (2023) holds 6 above a mean brightness of 180. Every one, and as many ordinary photographs (100 to 140) from the same cameras, was tested the same way:

Reference setPhotographsOwn camera matchedOther cameras matched
VISION (2017)32 above 180 indicative (n < 100)81.25 % [53.33-97.56] (26/32)0.00 % [0.00-0.50] (0/736)
VISION (2017)32 from 100 to 140 (same cameras) indicative (n < 100)87.50 % [68.18-100.00] (28/32)0.00 % [0.00-0.50] (0/736)
FloreView (2023)6 above 180 indicative (n < 100)83.33 % [35.88-100.00] (5/6)0.00 % [0.00-1.90] (0/192)
FloreView (2023)6 from 100 to 140 (same cameras) indicative (n < 100)83.33 % [35.88-100.00] (5/6)0.00 % [0.00-1.90] (0/192)

No other camera matched a bright photograph. On VISION (2017), which holds the most, a camera matched 81.2 % of its own bright photographs against 87.5 % of its ordinary ones. As with dark photographs, this is a measurement of the photographs these sets hold, not a rate for over-exposed photographs as a class.

The camera check when a picture is changed (the Attack Lab)

Ten ordinary photographs from each of the nine validation cameras of VISION (2017), as the camera saved them and changed in fourteen ways: saved again as JPEG from quality 95 down to 25 (the re-save keeps no camera metadata, as most programs and platforms do), reduced to 75, 50 and 25 %, and turned by 0.5, 1 and 5 degrees. Each picture was tested against every one of the nine cameras' whole-frame profiles, as the Vault tests a picture.

PictureOwn camera matchedOwn camera, the othersOther cameras matched
As the camera saved it98.9 % [94.0-100.0] (89/90)1 NO MATCH0 of 720
Saved again as JPEG, quality 9598.9 % [94.0-100.0] (89/90)1 NOT TESTED0 of 720
Saved again as JPEG, quality 8598.9 % [94.0-100.0] (89/90)1 NOT TESTED0 of 720
Saved again as JPEG, quality 7598.9 % [94.0-100.0] (89/90)1 NOT TESTED0 of 720
Saved again as JPEG, quality 6598.9 % [94.0-100.0] (89/90)1 NOT TESTED0 of 720
Saved again as JPEG, quality 5598.9 % [94.0-100.0] (89/90)1 NOT TESTED0 of 720
Saved again as JPEG, quality 4598.9 % [94.0-100.0] (89/90)1 NOT TESTED0 of 720
Saved again as JPEG, quality 3598.9 % [94.0-100.0] (89/90)1 NOT TESTED0 of 720
Saved again as JPEG, quality 2587.8 % [75.6-97.8] (79/90)11 NOT TESTED0 of 720
Reduced to 75 %98.9 % [94.0-100.0] (89/90)1 NOT TESTED0 of 720
Reduced to 50 %98.9 % [94.0-100.0] (89/90)1 NOT TESTED0 of 720
Reduced to 25 %78.9 % [56.7-95.6] (71/90)13 NOT TESTED, 6 INCONCLUSIVE0 of 720
Turned 0.5°1.1 % [0.0-6.0] (1/90)89 NOT TESTED0 of 720
Turned 1°0.0 % [0.0-4.0] (0/90)90 NOT TESTED0 of 720
Turned 5°0.0 % [0.0-4.0] (0/90)88 NOT TESTED, 2 INCONCLUSIVE0 of 720

No other camera matched any picture, changed or not.

As the camera saved them, 1 of the 90 photographs scored below the decision threshold against their own camera (NO MATCH), the kind of miss the rates above count. Where a changed picture's own camera did not match, the check never said NO MATCH. It said NOT TESTED where it could see the picture had been re-encoded or it could not tell what had been done to it: a miss there is not evidence of another camera. It said INCONCLUSIVE where a signature correlation score above the decision threshold was carried mostly by the finest detail, as another camera's picture enlarged after it was taken scores, so it could not single out this camera.

Ten photographs a camera make each cell indicative (n < 100): the curve shows where the check stops following a change, not a rate for any one change.

Compression history (Stage 2, ADR 0027)

The Compression History row says DOUBLE when a JPEG carries the trace of an earlier JPEG save. That is history, never evidence of editing: every messaging and social platform saves a picture again. Measured on 100 photographs from the training and validation cameras of each of VISION (2017) and FloreView (2023), each saved again with a standard library encoder at qualities on the standard (IJG) scale; the intervals are a cluster bootstrap over the 54 source cameras.

Where the trace is read: a second save at a higher quality than the first. v0.1's target, at least 90 % of these files found while at most 1.5 % of files saved once are called DOUBLE, is met in every case measured:

HistoryFound DOUBLE
Quality 65, then quality 95100.0 % [96.4-100.0] (100/100)
Quality 65, then quality 8599.0 % [94.6-100.0] (99/100)
Quality 65, then quality 7598.0 % [93.0-100.0] (98/100)
Quality 75, then quality 95100.0 % [96.4-100.0] (100/100)
Quality 75, then quality 8599.0 % [94.6-100.0] (99/100)
Quality 85, then quality 95100.0 % [96.4-100.0] (100/100)
Quality 90, then quality 95100.0 % [96.4-100.0] (100/100)

Where it mostly is not: a second save at a lower quality than the first. A camera original saved again is this case, because a camera stores its picture more finely than most saves that follow. These rates are published and are not held to the target (ADR 0027):

HistoryFound DOUBLE
Quality 85, then quality 7581.0 % [71.9-88.7] (81/100)
Quality 90, then quality 8583.0 % [74.2-91.0] (83/100)
Quality 90, then quality 7515.0 % [7.4-23.5] (15/100)
Quality 95, then quality 8513.0 % [1.1-27.0] (13/100)
Quality 95, then quality 753.0 % [0.0-9.4] (3/100)
FloreView (2023) camera original, then quality 9546.0 % [26.2-64.2] (46/100)
FloreView (2023) camera original, then quality 8540.0 % [21.9-56.6] (40/100)
FloreView (2023) camera original, then quality 7522.0 % [9.4-36.4] (22/100)
VISION (2017) camera original, then quality 9529.0 % [9.1-49.1] (29/100)
VISION (2017) camera original, then quality 8514.0 % [0.0-30.2] (14/100)
VISION (2017) camera original, then quality 753.0 % [0.0-9.4] (3/100)

Over every camera original measured, a second save is found in 25.7 % [14.5-36.8] (154/600) of files. More than half of them (322 of 600) keep almost nothing to read: where a camera's step divides the later save's, the second save leaves no trace there, and a step of 1, the commonest in camera tables, divides every step; among those files 3 were found. A result that is not DOUBLE therefore says nothing either way: it is not evidence that a picture was saved only once.

Saved twice at the same quality, a file keeps no readable trace at all (quality 75 twice: 4 %, quality 85 twice: 14 %, quality 95 twice: 29 % found, which is rounding noise, not the reading).

False alarms: photographs saved once at one quality (downscaled first, so that no earlier save survives) are called DOUBLE in 1 of 200 at quality 95, 0 of 200 at quality 85, 0 of 200 at quality 75; camera originals as the phones wrote them, in 0 of 100 (FloreView (2023)) and 1 of 100 (VISION (2017)).

Which platform made the last JPEG save (ADR 0019)

The Redistribution Trace reads a JPEG's header, which describes its last save: the quantisation tables, the colour subsampling, the Huffman tables, the frame type, the order of its segments, the metadata it carries and its size. It compares that with the signatures platform encoders write, fitted on VISION's training devices. A match is reported as CONSISTENT WITH the platform, for the last hop only: it names candidates, never the sender, the account or any earlier hop, and a file that matches nothing is NO KNOWN SIGNATURE, which does not mean it was never shared. Every result shows when each signature was observed, and a signature older than 365 days makes the result INCONCLUSIVE, because platforms change their encoders without notice.

The table holds WhatsApp and Facebook, as observed on 2017-03-21. VISION's copies date from when the set was assembled (it was published in 2017), and the date is the latest capture date among the sampled photographs, a lower bound on when the copies were made. So today every match reads INCONCLUSIVE until the platforms' current encoders are measured. Instagram, Telegram and X have no signature yet; their cells will be published as indicative (about 50 founder spot-checks each, below the 100-trial floor) until the 2026 transcode set is evaluated.

Files and the signature they matched (current or older than the limit; clustered by camera):

FilesSplitMatched WhatsAppMatched FacebookMatched none
VISION's WhatsApp copiestraining726/750 (96.8 % [90.4–100.0])0/750 (0.0 % [0.0–0.5])24/750 (3.2 % [0.0–9.6])
VISION's WhatsApp copiesvalidation (held out)450/450 (100.0 % [99.2–100.0])0/450 (0.0 % [0.0–0.8])0/450 (0.0 % [0.0–0.8])
VISION's Facebook copies (high quality)training0/750 (0.0 % [0.0–0.5])750/750 (100.0 % [99.5–100.0])0/750 (0.0 % [0.0–0.5])
VISION's Facebook copies (high quality)validation (held out)0/450 (0.0 % [0.0–0.8])445/450 (98.9 % [96.7–100.0])5/450 (1.1 % [0.0–3.3])
VISION's Facebook copies (low quality)training0/750 (0.0 % [0.0–0.5])750/750 (100.0 % [99.5–100.0])0/750 (0.0 % [0.0–0.5])
VISION's Facebook copies (low quality)validation (held out)0/450 (0.0 % [0.0–0.8])450/450 (100.0 % [99.2–100.0])0/450 (0.0 % [0.0–0.8])
VISION's camera originalstraining0/750 (0.0 % [0.0–0.5])0/750 (0.0 % [0.0–0.5])750/750 (100.0 % [99.5–100.0])
VISION's camera originalsvalidation (held out)0/450 (0.0 % [0.0–0.8])0/450 (0.0 % [0.0–0.8])450/450 (100.0 % [99.2–100.0])

WhatsApp: found in 450/450 (100.0 % [99.2–100.0]) of its held-out re-encoded copies (726/726 (100.0 % [99.5–100.0]) in training), and claimed for 0/1350 (0.0 % [0.0–0.3]) of the other held-out files (0/2250 (0.0 % [0.0–0.2]) in training).

Facebook: found in 895/900 (99.4 % [98.3–100.0]) of its held-out re-encoded copies (1500/1500 (100.0 % [99.8–100.0]) in training), and claimed for 0/900 (0.0 % [0.0–0.4]) of the other held-out files (0/1500 (0.0 % [0.0–0.2]) in training).

24 training WhatsApp copies kept the camera's own encode, all from one camera (Apple iPad2): WhatsApp passed them on without re-encoding, so their last save is the camera's, and they read NO KNOWN SIGNATURE, as they should. The 5 held-out Facebook re-encodes the table misses come from one camera (Apple iPhone4), saved at quality 91–92, where the training copies showed 71–90.

What it cannot tell apart. WhatsApp's 2016 encoder is plain libjpeg at quality 92 with its default settings, so any program saving that way writes the same header. The same photographs re-encoded by plain libjpeg measure it: at WhatsApp's settings every file reads as WhatsApp, and changing any one of the settings the signature reads takes the file off it. The 1600 px row's 4 matches are photographs from one camera (Apple iPad2) no longer than 960 px, which neither setting resizes, so they are the first row's files. That is why the trace says "consistent with" and nothing stronger.

Plain libjpeg re-encode of VISION's photographsTraining: matched WhatsAppValidation: matched WhatsAppMatched Facebook (both)
quality 92, 4:2:0, baseline, standard tables, 1280 px, no metadata: WhatsApp's 2016 settings60/60 (100.0 % [94.0–100.0]) (indicative)36/36 (100.0 % [90.3–100.0]) (indicative)0/96
the same at 1600 px4/60 (6.7 % [0.0–20.0]) (indicative)0/36 (0.0 % [0.0–9.7]) (indicative)0/96
the same at quality 850/60 (0.0 % [0.0–6.0]) (indicative)0/36 (0.0 % [0.0–9.7]) (indicative)0/96
the same with 4:4:4 colour0/60 (0.0 % [0.0–6.0]) (indicative)0/36 (0.0 % [0.0–9.7]) (indicative)0/96
the same keeping the camera's Exif0/60 (0.0 % [0.0–6.0]) (indicative)0/36 (0.0 % [0.0–9.7]) (indicative)0/96
the same with optimised Huffman tables0/60 (0.0 % [0.0–6.0]) (indicative)0/36 (0.0 % [0.0–9.7]) (indicative)0/96
progressive, optimised, quality 80, 2048 px0/60 (0.0 % [0.0–6.0]) (indicative)0/36 (0.0 % [0.0–9.7]) (indicative)0/96
the same with an ICC profile0/60 (0.0 % [0.0–6.0]) (indicative)0/36 (0.0 % [0.0–9.7]) (indicative)0/96

The signature table ships inside the signed ruleset and is refreshed without an engine release. The Facebook signature was fitted on 1,500 training copies from 15 cameras and the WhatsApp one on 726. The fit read the training devices only, but a first survey had read the held-out devices' headers too, so for this table they are not an untouched test; the sealed set is, at v1.0. Rows marked indicative have fewer than 100 files.

Platform copies: FODB (Forchheim, 2021)

FODB, the Forchheim Image Database (B. Hadwiger and C. Riess, "The Forchheim Image Database for Camera Identification in the Wild", Multimedia Forensics in the Wild (MMForWild 2020), Milan, Italy, January 10–15, 2021), holds 23,106 photographs of 143 scenes by 27 phones, each as the camera wrote it and as the copies Facebook, Instagram, Telegram, Twitter (now X) and WhatsApp made of it: Instagram, Telegram and Twitter copies beside their originals, which VISION does not have. By the phones' own clocks its photographs were taken in 2020 (2,600 of the 2,711 originals read here, 2,599 of them in August), so its copies were made more than two years after VISION's. Its authors allowed us to publish error rates measured on it; the images are not redistributed. Nothing in the product is set on it: no threshold, cut or parameter is fitted or tuned on FODB, and the phones of its locked split stay sealed for the final evaluation. The tables below are its 12 training and 7 validation phones, reported alike because neither was used to set anything. Both were read while the engine's handling of platform copies was designed (October 2026), so neither is an untouched test; the sealed phones are, at the final evaluation.

The camera check, as the Vault runs it. FODB holds no flat-field frames, so each phone's profile is made as the Vault makes one from natural photographs: 25 of the phone's own originals of scenes 1 to 30, the whole frame, of which the engine accepted 13 to 25. Each profile is then tested with the phone's pictures of every later scene, in all six versions, each picture handed to the engine whole with its own file facts, as the Vault's test hands it over, so no test picture shows a scene its profile was made from; every other phone's pictures test it as another camera. A profile made from natural photographs carries less of the sensor's signature than one made from flat fields, so the right camera is found less often here than with VISION's flat-field profiles above.

VersionRight camera found: trainingRight camera found: validationAnother camera matched: trainingAnother camera matched: validation
Camera original1,209/1,353 (89.4 % [82.4–95.2])575/788 (73.0 % [43.6–93.3])3/14,330 (0.02 % [0.00–0.06])0/4,227 (0.00 % [0.00–0.09])
Facebook0/1,353 (0.0 % [0.0–0.3])0/788 (0.0 % [0.0–0.5])2/7,441 (0.03 % [0.00–0.10])0/1,563 (0.00 % [0.00–0.24])
Instagram107/1,353 (7.9 % [0.0–18.6])15/788 (1.9 % [0.0–5.0])7/14,883 (0.05 % [0.01–0.10])3/4,728 (0.06 % [0.00–0.19])
Telegram221/1,353 (16.3 % [1.5–37.3])77/788 (9.8 % [0.1–25.6])2/14,323 (0.01 % [0.00–0.05])0/4,728 (0.00 % [0.00–0.08])
Twitter (now X)721/1,353 (53.3 % [36.7–67.3])389/788 (49.4 % [23.4–72.7])6/10,648 (0.06 % [0.01–0.13])0/3,636 (0.00 % [0.00–0.10])
WhatsApp280/1,353 (20.7 % [5.2–39.5])37/788 (4.7 % [0.0–14.1])8/14,840 (0.05 % [0.00–0.13])0/4,059 (0.00 % [0.00–0.09])

Another camera matched means another phone's picture read MATCH against a profile, over the comparisons the engine scored. Two of the phones are units of one model (HUAWEI VNS-L31); against each other they read MATCH in 0 of 1,130 scored comparisons.

Where another camera matched. 31 comparisons in all, on 23 pictures of 4 scenes taken by 8 phones, every one a dark picture: the brightest of their originals has a mean brightness of 32 of 255. The mirror check (ADR 0051) withheld 53 more, would-be MATCHes the profile's mirror image, which no camera produces, nearly reproduced, and 7 of the right camera's. 22 were on files whose history the file itself cannot tell (no camera metadata, and quantisation tables the engine does not name), which the Workbench and the signed report mark as outside the standard operating conditions above (condition (d)). Of the other 9, 1 were on camera originals and 8 on copies the engine recognised as re-encoded. Read as the Workbench reads a picture's conditions, 2 of them were on pictures inside all four conditions (1 on camera originals) and 7 on 5 pictures outside them (conditions missed, by picture: the signal-to-noise ratio 5, the size 2). By route: 16 through a resize of the whole frame; 10 through the middle of the frame cut to the copy's shape and resized (the route Instagram's squares take); 5 at the camera's own scale. Over files of unknown history the right camera matched 614/5,624 (10.9 % [3.7–18.8]) of its comparisons and another camera 22/41,914 (0.05 % [0.02–0.10]) of those scored.

So on very dark pictures a MATCH against a profile made from natural photographs can come from another camera, inside the conditions this page's rates apply to, though the engine measured here (version 0.7.23) withholds such a match when the profile's mirror image scores nearly as high (ADR 0051), which it did for 53 of the 84 it would otherwise report. Below a mean brightness of 50, a MATCH on a copy was about 9 times likelier to come from the right camera than from any one other camera, and on an original about 210 times (the table below).

By how dark the picture is. FODB holds 131 originals below a mean brightness of 50 (of 255) and 269 below 70 of the 2,711 read here, more than the reference sets of the dark-photograph section together. Each test picture is placed by its scene's original, training and validation together:

Original's mean brightnessTest pictures: originals, copiesRight camera found: originalsRight camera found: copiesAnother camera matched: originalsAnother camera matched: copies
below 50131, 65571/131 (54.2 % [34.8–74.5])31/655 (4.7 % [2.1–8.3])3/1,135 (0.26 % [0.00–0.77])28/5,094 (0.55 % [0.14–0.97])
50 to 70136, 68098/136 (72.1 % [59.7–82.8])44/680 (6.5 % [2.8–11.9])0/1,141 (0.00 % [0.00–0.32])0/5,074 (0.00 % [0.00–0.07])
70 and above1,874, 9,3701,615/1,874 (86.2 % [73.7–94.3])1,772/9,370 (18.9 % [11.8–28.2])0/16,281 (0.00 % [0.00–0.02])0/70,681 (0.00 % [0.00–0.01])

Other cameras matched pictures below 50 only: none of the 93,177 comparisons of brighter pictures, originals or copies, read MATCH. As in the dark-photograph section above, this is a measurement of the pictures this set holds, not a rate for under-exposed photographs as a class; that section's photographs were tested against profiles made from flat fields, these against profiles made from natural photographs.

The right camera's pictures that were not found, training and validation together:

VersionNot tested: no size or shape the check followsNot tested: read as a screenshotTested, not foundWithheld (INCONCLUSIVE)NO MATCH
Camera original0068108181
Facebook1,186095320
Instagram502,01310
Telegram11201,712190
Twitter (now X)0536441540
WhatsApp0531,757140

A copy the engine tested and did not find reads NOT TESTED, never NO MATCH: a platform's re-encode can remove the signature whether or not the camera is the same, so the miss says nothing about which camera took the picture. NO MATCH is said only of a camera original. A picture is not tested at all when no size or shape its dimensions imply is one the check follows (a copy smaller than a quarter of the camera's frame, or of a shape no whole-frame resize or centred crop explains), or when it reads as a screenshot. Among the camera originals, an original of another shape or size than the profile's frame is tried as a crop of that frame at the camera's own scale, and a miss there reads NOT TESTED (51); an original at the frame's size that carries a resampling trace reads NOT TESTED, since a miss on it cannot be told from a miss on an enlargement (11); an original the engine classes as a file of unknown history is tested as a copy is, and a miss reads NOT TESTED (6). Of the answers withheld here, 7 are the mirror check (ADR 0051), every one on a copy: would-be MATCHes the profile's mirror image, which no camera produces, nearly reproduced; the other 191 are the shared-pattern rule above: the profile carries a pattern other phones of its maker share, so a match on it could be any of them.

The Redistribution Trace on FODB. Its table holds WhatsApp's and Facebook's signatures as VISION's copies show them. Over every file of these phones it named a platform for 0 of 10,278 training files and 0 of 5,988 validation files. Every copy, of every platform, and every original reads NO KNOWN SIGNATURE: it names no platform wrongly, and finds none of FODB's WhatsApp or Facebook copies either, because their last saves differ from VISION's (below). That is why each result shows when its signature was observed, and why a match older than the table's limit reads INCONCLUSIVE: platforms change their encoders without notice.

VersionFilesFrameHuffman tablesQuantisation tablesLong side (pixels)Metadata segments
Camera original2,711baseline (2,711)standard (2,711)IJG quality 90–97 (1,286), no IJG quality (1,425)4,032 (714), 4,160 (516), 3,264 (286), 1 moreExif (816), JFIF Exif (611), 6 more
Facebook2,711progressive (2,711)optimised (2,711)IJG quality 71–86 (285), no IJG quality (2,426)960 (2,711)JFIF Photoshop (2,711)
Instagram2,711progressive (2,711)optimised (2,711)IJG quality 80 (428), no IJG quality (2,283)1,080 (2,569), 972 (142)JFIF Photoshop (2,711)
Telegram2,711baseline (2,288), progressive (423)optimised (1,856), standard (855)IJG quality 78–87 (2,711)1,280 (2,569), 854 (142)JFIF (2,568), JFIF ICC (143)
Twitter (now X)2,711progressive (2,711)optimised (2,711)IJG quality 85 (1,825), no IJG quality (886)2,048 (2,052), 2,080 (516), 1,600 (143)JFIF (2,052), JFIF ICC (659)
WhatsApp2,711progressive (2,711)optimised (2,711)IJG quality 75 (570), no IJG quality (2,141)1,600 (2,279), 1,024 (147), 1,599 (143), 1 moreJFIF (2,711)

Sources, licences and signing keys

The numbers on this page were measured on these reference sets, each used under its licence. The camera sets' licensors also gave their written permission to derive statistics, set thresholds and publish measurements with attribution:

The inpainting table's sets, whose licence allows these measurements with attribution:

The platform-copies set, whose authors allowed us to publish error rates measured on it:

The reference data each release ships is signed on hardware tokens, and an engine opens it only when it is signed under these roots. The two roots and the purpose keys were generated on two hardware tokens at the key ceremony of 2026-10-05 and never leave them; each purpose key carries a certificate from both roots, valid for two years.

KeyFingerprint (BLAKE3 of the Ed25519 public key)
Root, current (token A)a10a7507403b3fb335077d8eabc198cf0388efbdbcb5cf6ab0c6dceb923d3ae7
Root, emergency (token B)b04419274e35434ac3365797c81d764c7711c2d052ccd3fe5b725d4e9ca95ed8
Ruleset key1df605921143ef79f6c7da7108d90cf1dad59c0fb5d2c08a037dd34f449d50cd
Registry key334b85d63ef5e52f935bac7465facb05275783eda8ad6331afe5a02755a0e3b7
Licence keyb0340b4ac68bc60c0cd9e8b6fbd0e94e14004c942bdd5c6990f7501c54f77f99
Licence CA keyb44bb076775f2f261b6d001096052d3a2e573f0059c0e2c530efa5275b83f96c

To check a release: the key list the engine embeds must hold the two roots whose public keys hash to the root fingerprints above, and the ruleset's certificate must name the ruleset key above, signed by a root and in force on the ruleset's issue date. The open verifier publishes the same table.

Security advisories: one. Releases up to v0.7.23 open a .pvault profile with a decompression library (ruzstd 0.9.0) that a crafted file can make produce output without bound (RUSTSEC-2026-0344): importing such a file, with its passphrase, could stop the analysis worker in that browser tab. It could not read or send anything. This version carries the fixed library (0.9.1). A vulnerability is reported as /.well-known/security.txt describes.

What this version requests

The request allow-list for this version. The analysis runs in the browser, and the app requests only files from its own origin, all without a request body:

There are no analytics, no third-party scripts and no per-analysis beacon or completion counter, in any build. The app never sends file names, media hashes, dimensions, results or crash dumps; the diagnostic report is copied to the clipboard by the reader and never sent. The content security policy refuses every other origin, and the tests refuse a build that requests one. The timestamp relay, when it exists, will receive only a report's hash, and only when the reader asks for a timestamp. The host keeps its own request logs (address, path, time), as its privacy statement describes; no retention period is promised here. Every page shows an egress meter: an informational count of request bodies sent and of requests to other sites, not a security control.

What these numbers do not claim

Generated by tools/datasets/rates_page.py from ruleset R0 (provisional).

Analysis engine build (SHA-256 of core.wasm): 321de5e40d00314e473d0446fc460dbfd6abbf908f7149119825759a5cf63479