acceptodds
Under review as a conference paper at ICLR 2027

Anonymous Image Processing-Chain Forensics by Example

Abstract

Recovering an image’s processing history is a central problem in digital image forensics. Existing methods recover named operation sequences or compare processing traces. When operation names and configurations are unavailable, a few before-and-after examples provide another source of evidence. We introduce anonymous image processing-chain inference, a new task for ranking candidate histories between two image versions using these examples. Our central finding is that a history decision can be determined before its operations are fully identified. We derive sufficient conditions under which all example-compatible models in a stated parameter neighborhood prefer the same history, accounting for repeated operations and propagated estimation errors. We propose ChainPrint, which estimates reusable operation functions from paired examples within a declared model library and composes them to address diverse and compound history-inference problems without task-specific retraining. Shuffling example pairings degrades recovery; replacing sequential composition with additive operation changes lowers eight-step AI Top-1 accuracy by 41.4 percentage points. Across three image sources, eight-step Top-1 reaches 100.0% for conventional histories and 77.8% for AI-containing histories, with eight examples per operation and sixteen candidates. A fully enumerated two-dimensional control shows that ranking stability depends on which parameter directions the examples constrain, even when example fit and overall operation disagreement are matched. These results focus history recovery on the operation knowledge needed to distinguish competing histories, making example-based inference possible without requiring complete operation identification.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
digital image forensics, image processing history, operation estimation, compositional inference, image provenance

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_VFYTOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work addresses image processing chain forensics: given two versions of the same image (before and after), several input-output sample pairs for each unnamed operation, and a set of candidate processing histories, the task is to rank these histories. ChainPrint fits executable functions for each operation using a predefined model library (parameterized standard operators plus eight SD-1.5 inpainting configurations), chains them together according to the candidate histories, and ranks the results based on the final Mean Squared Error (MSE). Top-1 accuracy over eight steps is 100% for standard processing chains and 77.8% for chains involving AI-based editing.

Critical strengths
  1. The task is clearly defined, and a single framework covers various types of forensic problems—specifically those involving sequence, repetition, and partial spans.
  2. The reporting is transparent. Failure counts are tallied, pairwise confidence intervals are provided, and the data in the supplementary CSV files aligns perfectly with Tables 1–4.
Critical weaknesses
  1. My primary concern is that the model library already encompasses the exact processes used to generate the data. The AI ​​editor's prior corresponds precisely to the eight configurations sampled by the test editor (App. D, App. I); the paper itself acknowledges that it "does not discover an unknown editor." Regarding standard operators, the search spaces for JPEG and resampling cover the exact Pillow settings used during generation (App. F, App. I). While generic filters are used for blurring and sharpening—and the out-of-library NumPy experiment (Fig. 3c, 99.4%) offers some support—this applies only to standard operators. Thus, the 100% success rate is unsurprising. The 77.8% figure also aligns closely with the proportion of renderable candidates (140/180), given that a complete ranking implies correctness ("correct whenever complete," Sec. 5.2).

  2. I fail to see a significant methodological contribution in the comparative experiments. Several modified baselines performed worse than random chance (1/16) even after up to 420 updates, which is inconclusive. The only direct comparison—using the same prior for joint fitting—showed essentially equal performance (90.32 vs. 89.68, with all confidence intervals including zero). In Table 3, correction, in-sample selection, and refinement shifted the Top-1 accuracy by at most 0.8 percentage points, with confidence intervals again spanning zero. Significant performance drops occurred only when replacing sequential composition with an additive model (−41.4) or shuffling pairings—both of which serve as weak baselines. The remaining benefit is a 5.3× reduction in preparation time, measured across six sources.

  3. There is a fundamental disconnect between the theory and the experiments. The paper states that the exact invariance condition (Prop. B.2) does not assume validity for JPEG or AI editing operators, nor does the estimator compute stability or recovery certificates (Sec. 4.3). The supporting evidence relies on a 2D comparison, consistency across a maximum of nine standard model sets, and four configurations of a single editing operation.

Questions for authors
  1. How does ChainPrint behave if the AI ​​edit falls outside the defined prior—such as with an unseen prompt, mask, or seed, or when switching to a different inpainting model?
  2. In the two-step process, errors persist even when the ranking is complete (FiveK: 164/240 complete, Top-1 67.1%). What is the source of these errors? Additionally, why did the eight-step FiveK process only complete 34/60 cases and generate just 58 queries?
  3. Aside from setup time and the small-scale 12-source AI Recall@5 diagnostic (91.7 vs. 66.7), are there other scenarios where ChainPrint outperforms joint fitting?
6 comments
AuthorsOpenReview ↗

Response to VFYT (1/4) — Library Coverage and Unseen AI Edits (W1, Q1)

W1 (excerpt): My primary concern is that the model library already encompasses the exact processes used to generate the data.

Q1 (summary): How does ChainPrint perform with an unseen prompt, mask, or seed, or with a different inpainting model?

Thank you for raising this concern. The original benchmark does cover its generation settings. The prior supplies a range of functions to fit, and the anonymous pairs supply the observed behavior of each operation. Our view is that history recovery needs an approximation that preserves differences between candidate histories. We added experiments with generating implementations absent from the library to test this view.

W1 · A new implementation under the fixed library

We replaced Pillow JPEG with FFmpeg's native MJPEG encoder, keeping the function library fixed. Both implementation and operating points changed; this was not a quality-matched encoder comparison. Refitting with eight anonymous pairs per operation raised recovery from 128/180 to 172/180 on the same FFmpeg queries (Table R1a). The examples were separate from the queries.

Table R1a. Fixed original library, including Pillow JPEG but no FFmpeg implementation. Both rows use the same 180 FFmpeg eight-step queries, 16 candidates and MSE. Only the support set changes.

Support source, 8 pairs/operationAll 180With JPEG, 156Without JPEG, 24
Original, using Pillow JPEG128/180104/15624/24
Refreshed, using FFmpeg JPEG172/180148/15624/24

All 44 net gains came from JPEG-containing histories. Generator names describe the experiment; fitting received anonymous image pairs, not operation-type or encoder labels.

Q1 · Unseen AI configurations and editors

Refreshing examples was insufficient for the original eight fixed AI configurations (Table R1b). Complete rankings could still select the wrong history.

Table R1b. Original prior8 (eight fixed SD1.5 AI configurations) under four-step mismatch, with 16 candidates and MSE. K is the number of accepted support pairs per operation actually used.

Change to the generating editorKPlannedTargets generatedComplete rankingsCorrect
SD1.5, seed only63631314/36
SD1.5, prompt + mask + seed612991/12
SD1.4-based masked editor83635272/36

The first two studies selected K=6 after checker refusals, using different support subsets. Their query sources overlap. Prompt and mask changed jointly; these rows do not isolate each factor or compare shift severity.

For SD1.4, we therefore extended the fitted-function family. The estimator was not given the SD1.4 implementation; it learned a pixelwise RGB affine approximation from the anonymous pairs. We froze the approximation before generating the 24 additional two-step targets. With the same eight SD1.4 support pairs, candidates and MSE, recovery rose from 5/24 to 14/24 (Table R1c).

Table R1c. Same 24-source SD1.4 two-step cohort, K=8 and 16 candidates, pixel MSE; the other three fitted operations stay fixed. SD1.4 generates examples and targets but is absent from the estimator.

EstimationTargets generatedComplete rankingsCorrect / planned
Original prior823/2417/245/24
Generic affine approximation23/2423/2414/24

On the 17 common-complete queries, correct counts changed from 5 to 9. Five of the six newly complete rankings were correct.

We also tested a constrained, shared-slope RGB affine map at each pixel, paired with local directional power scoring. After freezing both, we generated four-step queries from a separate 24-source cohort, disjoint from earlier mismatch queries and supports. Each true history called the same AI operation twice.

Table R1d. Separate SD1.4 four-step cohort: frozen shared-slope family, 16 candidates, all 24 targets and rankings complete. No original-prior8 comparison was run on this cohort.

AI support examplesEndpoint scoreCorrect / 24
8 natural-image pairsPixel MSE6/24
8 natural-image pairsLocal directional power12/24
16 natural-image pairsPixel MSE8/24
16 natural-image pairsLocal directional power10/24
8 natural + 8 already-edited-input pairsPixel MSE8/24
8 natural + 8 already-edited-input pairsLocal directional power16/24

With eight pairs, the extension recovered 12/24 histories (Table R1d). In the last support setting, changing the score gave +33.33 percentage points (paired 95% interval [8.33, 58.33]). These results support learning an approximation for history recovery when the generating editor is absent from the available functions. We will add the extension and its evaluation alongside the original-prior mismatch results.

AuthorsOpenReview ↗

Response to VFYT (2/4): Ranking Errors and Execution Availability (W1, Q2)

W1 (excerpt): The 77.8% figure also aligns closely with the proportion of renderable candidates (140/180), given that a complete ranking implies correctness ("correct whenever complete," Sec. 5.2).

Q2 (summary): What causes errors among complete two-step rankings, and why does FiveK produce only 58 eight-step targets and 34 complete rankings out of 60 planned queries?

Thank you for raising this issue. Your reading of the eight-step 77.8% result is correct. Of 180 planned queries, 173 produced targets and 140 yielded complete rankings, all correct. In this experiment, 77.8% therefore also equals the complete-ranking rate. We have traced the errors among complete two-step rankings to operator estimation and corrected them.

Errors among complete rankings (Q2)

FiveK's 67.1% is 161/240 overall, with three errors among its 164 complete rankings. We traced every error among complete two-step rankings, three from FiveK and one from COCO. Each reversed blur→AI into AI→blur, although the selected AI configuration was correct. Small blur-approximation errors changed the order comparison after the subsequent AI edit.

We therefore improved parameter fitting. The operator library and eight anonymous support pairs per operation stayed unchanged. We searched parameters using the existing models' hard-quantized forward passes. Adoption depended only on content-excluded support error. For the relevant operation, held-support MSE8 fell from 0.109775 to 0.0001443.

Table R2a. Original versus improved fitting, with the same query targets, candidates, MSE and complete-ranking criterion within each cohort; K=8 throughout. All counts retain planned queries.

Two-step cohortFitterTargets generatedComplete rankingsCorrect / planned
Original 720 queriesOriginal697/720604/720600/720
Original 720 queriesImproved697/720608/720608/720
FiveK 240-query subsetOriginal225/240164/240161/240
FiveK 240-query subsetImproved225/240164/240164/240
Separate new 24-source SD1.5 cohortOriginal23/2419/2418/24
Separate new 24-source SD1.5 cohortImproved23/2419/2419/24

Table R2a shows 600/720→608/720, with all four original complete-ranking errors repaired and no regression among the original 600 correct decisions. Another four queries, all conditions of one COCO source, gained complete, correct rankings. The 720 queries comprise 180 sources with four conditions each; FiveK is a subset, not another independent test.

The new 24-source test used a fixed blur/SD1.5 configuration, distinct from the SD1.4 cohorts in Response 1. Models were frozen before target generation. The improved fit corrected another instance of the same order confusion, again without reversing any previously correct decisions.

FiveK's 58 eight-step targets and 34 complete rankings (Q2)

Table R2b. Original eight-step availability and recovery. The FiveK row is included in the all-dataset row; these results do not use the new fitting.

CohortPlannedTargets generatedComplete rankingsCorrect / planned
All datasets180173140140/180
FiveK60583434/60

For FiveK, two of 60 target generations were refused by the editor's checker, leaving 58. Among these, 24 queries lacked candidate scores following execution refusals, leaving 34 complete, correct rankings (Table R2b). We will report all three counts separately so both failure points are visible.

For missing scores, we also added a conservative two-step decision rule. Errors on known pixels outside the mask lower-bound a missing candidate's MSE. A winner is returned only when these bounds exclude the missing candidates.

Table R2c. Different decision rules on the same original 720 two-step queries, using the improved estimates.

RuleDecisions / 720Correct histories / 720
Require complete rankings608/720608/720
Conservative lower-bound rule680/720680/720
Rank available candidates only697/720697/720

The conservative rule yielded 680/720 decisions, all correct, while direct available-only ranking yielded 697/720 correct histories (Table R2c). The bounds show why missing candidates cannot match or beat the winner under the fitted models. We will report these partial decisions separately from complete rankings.

AuthorsOpenReview ↗

Response to VFYT (3/4): Same-Prior Joint Comparisons and Module Contributions (W2, Q3)

W2 (summary): The learned baselines provide weak evidence, same-prior Joint performs similarly, and the selection, correction and refinement gains in Table 3 are small, with intervals including zero.

Q3 (summary): Beyond preparation time and the small 12-source AI Recall@5 diagnostic, does ChainPrint outperform Joint elsewhere?

Thank you for identifying the gaps in our strong-baseline comparisons and module evidence. The added experiments give two direct answers: ChainPrint outperforms Joint with matched native adaptation on real platform chains (Q3), and refinement and correction improve history-recovery accuracy under implementation mismatch (W2).

Same-prior Joint comparisons (W2, Q3)

We used public historical chains produced by actual platform uploads and downloads, fixing 52 query sources from 26 cameras absent from the support set. The method and split were frozen before query acquisition.

Table R3a. Matched native adaptation (P2), same 52 sources. Both methods share supports, prior budget, native dimensions and geometry, candidates and MSE. Each uses three separate K=8, G=1 cold fits, then assembles the operator library.

MethodTwo-step accuracyThree-step accuracy
Native-adapted Joint41.45%13.46%
Native-adapted ChainPrint47.44%16.74%
CP−Joint, pp (95% interval)+5.98 [2.56, 9.19]+3.28 [1.85, 4.84]

Table R3a shows gains of 5.98 and 3.28 percentage points. Intervals are clustered by the 26 cameras and correspond to prespecified secondary comparisons without multiplicity correction. This extends the evidence for Q3 to real-platform Top-1 recovery, with both methods receiving the same information and prior budget.

The advantage concerns this native-adaptation protocol, not every possible native Joint design. The original pixel version did not outperform Joint, and the JPEG and cross-checkpoint comparisons did not establish a consistent advantage. We will retain these results.

Roles of refinement and correction (W2)

Your observation about the small gains in the original Table 3 is correct. To test whether the modules improve history decisions, we fixed supports, prior, selection procedure, queries and candidates under FFmpeg JPEG implementation mismatch, changing only refinement and correction.

Table R3b. Same refreshed FFmpeg support set (8 pairs/operation), original function library, 180 eight-step queries, 16 candidates and MSE. Only the two module switches change.

RefinementCorrectionCorrect / 180
Off (initial estimates)Off139/180
Off (initial estimates)On139/180
OnOff161/180
OnOn172/180

The full configuration recovered 172/180 histories, a net 33 more than no-refinement (+18.33 percentage points) and 11 more than no-correction (+6.11 points), as shown in Table R3b. Source-paired 95% intervals were [12.78, 23.89] and [2.78, 9.44]. All gains came from JPEG-containing histories; the 24 JPEG-free histories remained correct.

The ablation also reveals an interaction in this condition: correction gives no gain on the initial estimates, but improves accuracy after refinement. We will report the original small-gain results alongside this mismatch experiment.

Preparation cost was also independently checked.

Table R3c. Separate seven-condition conventional-processing experiment, original frozen fitters, same host and four CPU cores. Preparation CPU includes common AI-cache construction.

Cost comparisonObservation
Joint/CP preparation CPU ratio5.66–5.90
Conditions with slower CP queries5/7

The 5.66–5.90 Joint/CP preparation CPU ratio in Table R3c concerns building an operator library from new anonymous examples. It does not apply to the new parameter search in Response 2 or the P2 adaptation above. Query costs will be listed separately, since ChainPrint queries were slower in five conditions.

We agree that the below-chance learned-baseline results do not establish an advantage for our estimator. We will make same-prior Joint and these controlled ablations the principal experimental evidence, with learned baselines retained as supplementary results.

Reviewer_Bup9OpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies anonymous image processing chain inference. The operation names and their configurations are unknown, and only a few before and after pairs are available. Given two image versions, the task is to rank candidate processing histories. The authors propose ChainPrint. It estimates reusable operation functions and composes them in candidate order, so one framework handles long chains, operation order, repeated operations, and configuration differences. The core insight is that deciding the correct history may require less operation knowledge than full operation identification. The authors characterize this gap in theory, then support it with direction controlled experiments, a real editor instance, and experiments on three image sources.

Critical strengths
  1. The problem setting has value. Describing an unknown operation through examples gives image forensics a clear new angle.
  2. The method is intuitive. One set of operation models serves different history inference problems, and it needs no refitting.
  3. The theoretical analysis targets the right question: which operation differences actually change the history decision. The direction controlled experiments support this point well.
  4. Long chain recovery is strong. The pair shuffling and composition ablations show clearly why the method works.
Critical weaknesses
  1. The experiments rely on a preset model library and a fixed AI editor, so applicability to operations outside the library remains unclear.
  2. The theoretical stability conditions have not been turned into a usable confidence measure, and validation on the real editor stays limited.
  3. Against joint fitting under the same prior, the main advantage is preparation cost. The accuracy gain from the refinement and correction modules is not obvious.
Questions for authors

I lean toward acceptance. I am particularly impressed by the paper's decoupling of operation identification and history inference, supported by both theory and experiments. I have remaining questions and await the authors' rebuttal.

  1. How does the method usually fail when a real operation is absent from the model library?
  2. Does performance stay stable when one chain contains several AI edits?
  3. How much does slight image misalignment affect the ranking?
  4. Are the results stable when the operation examples are replaced with another set?
  5. Can the score gap between candidates provide a simple confidence measure?
  6. Which operations need the refinement and correction modules most?
  7. Do other editors or operation targets show the same pattern, where effects differ but the history decision does not?
3 comments
AuthorsOpenReview ↗

Response to Bup9 (1/4) — New Operations and Repeated AI Edits (W1, Q1, Q2)

W1: The experiments rely on a preset model library and a fixed AI editor, so applicability to operations outside the library remains unclear.

Q1: How does the method usually fail when a real operation is absent from the model library?

Q2: Does performance stay stable when one chain contains several AI edits?

Thank you for your thoughtful comments. Among the questions you raise, we find Q1 and the related concern in W1 particularly important: how should our method operate when the real operation is absent from the library, or difficult to describe accurately with the available functions? This question gets to the heart of the method's applicability. We would first like to explain our view, then present the experiments we added to investigate it.

W1 / Q1 · When the library misses an operation

History recovery needs an operation approximation that distinguishes candidate histories; it does not always require exact reconstruction of the operation. An unlisted implementation may therefore be handled by fitting anonymous examples. In a chain, approximation errors can accumulate or cancel; what matters is whether they change the loss differences between candidates. This gives a common way to analyze familiar and unseen operations, without assuming comparable error magnitudes.

We first replaced Pillow JPEG with FFmpeg's native MJPEG encoder. With the function library unchanged, refreshing examples improved recovery from 128/180 to 172/180 (Table R1a), without encoder identities or operation labels. For the 156 JPEG-containing histories, correct recoveries rose from 104 to 148; all 24 other histories remained correct. This tests a new implementation and operating points, rather than quality-matched encoders.

Table R1a. FFmpeg MJPEG test: eight-step histories, 16 candidates, eight pairs/operation. The original library contains Pillow JPEG and no FFmpeg executor; the same 180 queries are used in both rows.

Fitting settingJPEG examples fromCorrect/planned
Original examples; fixed libraryPillow JPEG128/180 (71.11%)
Refreshed examples; fixed libraryFFmpeg MJPEG172/180 (95.56%)

For the new AI editor, however, refreshing examples was insufficient. We replaced SD1.5 inpainting with an SD1.4-based masked editing pipeline. Even with eight pairs from this editor, the original prior produced 17/28 wrong complete two-step rankings and 25/27 wrong complete four-step rankings. A complete ranking can still be unreliable when the operation model fits poorly.

Table R1b. SD1.4 masked-editor test: eight anonymous pairs/operation, 16 candidates. Within each length, supports and queries are identical across priors; both lengths use the same 36 sources.

Available AI proposals2-step correct/planned4-step correct/planned
Original eight SD1.5 configurations11/36 (30.56%)2/36 (5.56%)
Original eight + actual SD1.4 editor35/36 (97.22%)30/36 (83.33%)

We added the actual new editor as a ninth proposal before query scoring. Every anonymous slot could select from the same expanded prior; we did not identify the editor's slot. The estimator selected it from the examples and recovered most decisions (Table R1b), using the same fitting and history-inference procedure. We supplied that proposal; automatic discovery of arbitrary missing editors remains untested. All accuracy denominators in Tables R1a–R1b retain failed target generation and incomplete rankings; the true anonymous history remains a candidate.

When the selected model cannot explain the support examples, we believe the system should warn or abstain. A support-only diagnostic flagged this SD1.4 mismatch, but passing it does not guarantee a correct query decision. We discuss this distinction under W2/Q5.

Q2 · Recovery with repeated AI edits

Every four-step candidate in Table R1b contains two calls to the same AI operation. In an exploratory eight-step test, true histories contained four AI calls using two seed configurations of SD1.5 (Table R1c).

Table R1c. Separate eight-step repeated-AI test: four AI calls in each true history, two SD1.5 seed configurations.

Planned queriesTargets generatedRankings completeCorrect histories
24191313

All 13 complete rankings were correct, giving 13/24 (54.17%) overall. Five target-generation refusals and six incomplete rankings account for the other queries. These tests show recovery with repeated AI edits, with a substantial availability limit. Candidate histories also differ across lengths, so they do not isolate the effect of AI-call count.

We will add these experiments, distinguish fixed-library adaptation from proposal expansion, and report generation/completion counts alongside accuracy.

AuthorsOpenReview ↗

Response to Bup9 (3/4) — Stability, Confidence, and Other Editors (W2, Q5, Q7)

W2: The theoretical stability conditions have not been turned into a usable confidence measure, and validation on the real editor stays limited.

Q5: Can the score gap between candidates provide a simple confidence measure?

Q7: Do other editors or operation targets show the same pattern, where effects differ but the history decision does not?

W2 · Computing the stability check

Thank you for asking how the stability condition can guide an actual decision. We implemented a per-query check of Eq. (9) on four-step image chains. We fitted one continuous horizontal resampling parameter from eight support pairs; the other three operations were fixed and known. Support RMSE tolerances give a conservative outer interval ∣δ∣≤r|\delta|\le r covering compatible perturbations in the declared parameter neighborhood.

For nominal winner kk and rival hh, the MSE gap is exactly

mh(δ)=m^h+vhδ+ahδ2.m_h(\delta)=\widehat m_h+v_h\delta+a_h\delta^2.

The outer interval gives WS(vh)≤∣vh∣rW_S(v_h)\le|v_h|r. The uniform remainder bound is ρh=∣ah∣r2+10−8\rho_h=|a_h|r^2+10^{-8}, including a numerical allowance. Equation (9) therefore certifies stability when

m^h>∣vh∣r+ρhfor every h≠k.\widehat m_h>|v_h|r+\rho_h\quad\text{for every }h\ne k.

At tolerance 1/255, this check certified 180/180 decisions (Table R3a). We also minimized the quadratic gaps over the same outer interval as a reference.

Table R3a. Controlled image resampling: four steps, eight fixed support pairs, one unknown continuous parameter and three known operations; 180 queries, 16 candidates each. Counts concern stability, not estimator accuracy.

Per-support RMSE toleranceOuter radius rrEq. (9) certifiedQuadratic minimum positive
1/2550.124869180/180180/180
2/2550.24973899/180179/180
4/2550.49000049/180121/180

Q5 · Can the gap serve as confidence?

The relative gap was useful for screening. The raw gap was weak, and frozen thresholds fell short of the 95% development accuracy target. A history can have a large lead even when it fits the observation poorly. We compared raw gap and negative minimum loss with relative gap, which divides the gap by the sum of the two lowest losses plus 10−1210^{-12}, and a logistic score combining minimum loss and raw gap on a log scale.

On 90 development sources, we chose the highest-coverage threshold achieving at least 95% accuracy and retaining at least 10 distinct sources, then froze the model and thresholds. On 90 held-out sources under the same shifts, relative gap retained 429/810 cases: 397/429 (92.54%) were correct, versus 475/810 (58.64%) before screening.

Table R3b. Frozen screening rules on ImageMagick: refreshed operation models, original pixel scoring, 24 additional FiveK contents × 11 shifts = 264 records. Thresholds come from the separate 90-source development set. All rows use the same 264 records.

Screening ruleAUCRetained/264Correct/retained
No screening—264/26495/264 (35.98%)
Raw gap0.555No qualifying threshold—
Negative minimum loss0.66785/26442/85 (49.41%)
Relative gap0.74926/26420/26 (76.92%)
Two-score logistic0.773127/26471/127 (55.91%)

Table R3b shows both the gain in retained accuracy and its coverage cost: relative gap retained only 26/264 cases. These are empirical screening scores, not calibrated probabilities or Eq. (9) certificates. The support-fit check asks whether the operation model explains the examples: at tolerance 2/255 it detected the original-prior SD1.4 mismatch, but the refreshed ImageMagick model passed despite three aligned-query errors.

Q7 · Does the pattern hold for another editor?

Yes. With the SD1.4-based masked editor, we fixed eight support pairs, six observed queries and 16 candidates per query, varying only guidance. Settings 7.25 and 7.5 both met per-support RMSE tolerance 2/255.

Table R3c. Fixed-support, fixed-query SD1.4 check: guidance 7.25 versus 7.5, eight support pairs and 16 candidates/query; both satisfy tolerance 2/255. Direct-edit RMSE measures the output change from guidance 7.5 on the same input.

Planned queriesComplete under bothSame uniquely correct historyDirect-edit RMSE >1/255 and stable
6552

The five jointly complete queries retained the same uniquely correct history (Table R3c). Two, from Flickr8k and COCO, also had direct-edit RMSE above 1/255. A candidate-execution refusal left the sixth ranking incomplete; the observed targets remained fixed.

We will report the stability check, frozen screening accuracy and coverage, and second-editor results in the revision.

The authors' response has fully addressed my remaining concerns. Overall, I consider this paper to make novel and well-supported contributions. Therefore, I maintain my rating of 4.

Reviewer_nkzWOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces anonymous processing-chain inference: ranking candidate operation histories between two images using a few before/after examples per unnamed operation. ChainPrint fits one function per operation from a declared library, composes them, and scores candidates by MSE, with sufficient conditions for when the ranking is already determined.

Critical strengths

I like that the task is posed cleanly and the controls are careful. Shuffling the example pairings costs 71.8 points and swapping sequential for additive composition costs 41.4 (Table 3), which convinces me the examples are doing real work and not just the candidate prior.

The second thing I value is the enumerated control in Fig. 4a-c (24/32 vs 0/32 depending on the direction of the constraint). It makes the directional-uncertainty argument concrete, even if only in a toy setting, and it is the part of the paper I learned the most from.

Critical weaknesses

My main worry is that the benchmark mostly measures search inside a library that already contains the process that generated the data (App. A, App. D). "Anonymous" then largely means parameter identification within a known family. Accuracy is already saturated at K=2 (Fig. 3a), the deep baselines were designed for other tasks, and the one comparator that seems fair, joint fitting with the same prior, is statistically tied with ChainPrint (89.68 vs 90.32, Sec. 5.2). Without an out-of-library operation or a real pipeline, the C2PA and social-sharing motivation is not tested.

The theory also feels disconnected from the method. Eqs. 6-7 and 9 need orthogonality, commuting maps and a remainder bound, and I could not find any check of these for JPEG, resampling or SD inpainting. I did not go through the appendix proofs line by line, so the derivations may well be fine; my concern is applicability, especially since refinement and correction change little in Table 3.

Questions for authors

What happens when the true operation is not in the library, for example a different JPEG encoder or inpainting model? This matters most for my score.

Can the conditions in Eq. 6 or Eq. 9 be certified for even one real operation, or are they only illustrative?

Have you tried any real pipeline, such as platform recompression? A short note on whether pixel-aligned inputs are required would help.

5 comments
AuthorsOpenReview ↗

Response to nkzW (1/4): Unseen Operations and Anonymous Examples (W1, Q1)

W1 (summary): The benchmark largely measures search within a library that contains the generating processes. Anonymity largely amounts to parameter identification within known families, and accuracy already saturates at K=2.

Q1 (summary): What happens if the library lacks the actual operation needed, such as a different JPEG encoder or inpainting model? This point is crucial to my score.

Thank you for raising this central question. We prioritized Q1 and tested both cases. JPEG adapts through refreshed examples within the fixed library. For the tested AI mismatch, adding learnable approximations improved recovery without providing the actual editor.

For JPEG, Table R1a compares the same new FFmpeg queries. The baseline retains fits from the original Pillow examples; the second row refits using examples from the new encoder. Recovery improves from 128/180 to 172/180, with the library and test queries unchanged.

Table R1a. JPEG implementation mismatch: eight steps, the same 180 queries, 16 candidates per query, and eight anonymous pairs per operation.

SettingKnown (JPEG library)Unknown (test generator)JPEG examplesCorrect / 180
BaselinePillow JPEG; no FFmpegFFmpeg MJPEGOriginal Pillow examples128/180
RefittedPillow JPEG; no FFmpegFFmpeg MJPEGRefreshed FFmpeg examples172/180

The 156 JPEG-containing histories improve from 104/156 to 148/156; the 24 without JPEG remain 24/24. All 44 corrections involve JPEG. Both implementation and operating points changed; the encoders were not quality-matched.

For AI, Table R1b compares function families using the same examples. The original library includes eight SD1.5 configurations. We added per-pixel affine and residual functions fitted from the same eight anonymous pairs, selecting models by leave-one-support-out error. Neither row supplies the actual SD1.4 editor. Other operations, candidates and pixel-MSE scoring stay unchanged. Parameters and rules were frozen before confirmation on 24 sources separate from the supports and earlier mismatch queries. Recovery improves from 5/24 to 14/24.

Table R1b. SD1.4 two-step confirmation: the same 24 planned queries, eight AI pairs, 16 candidates and MSE scoring. Both rows use the same 23 generated targets.

SettingKnown (AI library)Unknown (test generator)AI examplesComplete / 24Correct / 24
BaselineSD1.5 prior; no SD1.4SD1.4 masked editorEight SD1.4 pairs17/245/24
Learned approximationSD1.5 prior + affine/residual families; no SD1.4SD1.4 masked editorSame eight SD1.4 pairs23/2414/24

Generator names describe the experiment; the estimator receives no identity or operation-type labels. On the 17 jointly complete queries, recovery also improves from 5/17 to 9/17. Gains include both decision changes and greater availability. Although 14/24 remains limited, in this SD1.4 setting, example-fitted substitutes support history recovery without supplying the actual generator to the estimator.

In practice, additional information can also supply an adequate model. A content producer may retain original and resulting images and provide calibration examples or partial workflow records. Knowing the tools does not determine their order. If a configuration-matched executable editor proposal is also available before inference, our separate 36-source SD1.4 comparison tests this condition: adding the matching editor, with supports and queries fixed, improves two- and four-step recovery from 11/36 and 2/36 to 35/36 and 30/36, respectively.

An application may only need to know whether AI editing preceded or followed resizing. That relation could remain consistent when the full history is ambiguous. We see partial-order decisions as a target for evaluation; current full-history accuracy does not measure their performance. Their value should be assessed against what the same workflow information alone supplies.

K=2 saturation concerns example requirements in the original matched setting, not adaptation outside it. We will add these experiments and distinguish fixed-library adaptation, learned approximations and recovery with additional editor information.

AuthorsOpenReview ↗

Response to nkzW (2/4): Testing the Conditions on Image Operations (W2, Q2)

W2 (summary of the theory concern): The theory and method seem disconnected. Equations (6)–(7) and (9) require orthogonality, commuting maps and remainder bounds, but there is no verification for JPEG, resampling or SD inpainting.

Q2 (summary): Can the conditions in Eq. (6) or Eq. (9) be checked for an actual operation, or are they only illustrative?

Thank you for focusing the question on applicability. Our direct answer to Q2 is yes: we have computed per-query stability certificates from Eq. (9) in a controlled continuous resampling experiment on actual images. We also checked compatible parameters for actual JPEG and used error-propagation analysis to repair AI order reversals.

The conditions have different roles. Equations (6)–(7) describe exact cancellation; preserving one winner does not require every relative loss to remain unchanged. Equation (9) tests whether operation changes allowed by the examples can reverse the winner. It does not require JPEG or AI to satisfy Eq. (7)'s orthogonal decomposition or commutativity.

1. Resampling: computing Eq. (9). With eight fixed support pairs, we tested four-step histories and 16 candidates on 180 images. The unknown operation was horizontal interpolation, Aθ(x)=(1−θ)x+θright1⁡(x)A_\theta(x)=(1-\theta)x+\theta\operatorname{right1}(x), with one unknown parameter and three known operations. For nominal winner kk, let m^h=ℓ^h−ℓ^k\widehat m_h=\widehat\ell_h-\widehat\ell_k and δ=θ−θ^\delta=\theta-\hat\theta. Each competitor margin is exactly mh(δ)=m^h+vhδ+ahδ2m_h(\delta)=\widehat m_h+v_h\delta+a_h\delta^2. Supports give a conservative outer radius rr containing the compatible set. We then check:

m^h>∣vh∣r+∣ah∣r2+10−8,h≠k.\widehat m_h>|v_h|r+|a_h|r^2+10^{-8},\qquad h\ne k.

Here WS(vh)≤∣vh∣rW_S(v_h)\le|v_h|r and the uniform remainder ρh=∣ah∣r2+10−8\rho_h=|a_h|r^2+10^{-8} covers the whole interval; 10−810^{-8} is a verified numerical guard.

Table R2a. Eq. (9) certificates: four-step horizontal interpolation, eight fixed support pairs, one unknown parameter and three known operations; 180 queries, 16 candidates each.

Per-example support RMSE tolerance, pixels in [0,1]Eq. (9) certified / 180
1/255180/180
2/25599/180
4/25549/180

Table R2a certifies all 180 queries at tolerance 1/255, throughout the outer interval. At tolerance 1/255, 14 queries also admit perturbations verified compatible with every support pair that change endpoint RMSE by more than 1/255, while remaining covered by the certificate. Operation effects can therefore vary while the history decision remains determined.

2. Actual JPEG: testing whether exact cancellation is necessary. With eight fixed support pairs and Pillow JPEG as the only unknown operation, we exhaustively evaluated qualities 40–100. At support tolerance 4/255, each cell had seven compatible qualities; the other three operations were known. Table R2b shows that 168/180 retain the correct winner across every compatible quality despite failing the Eq. (6) check.

Table R2b. Separate Pillow-JPEG finite-domain check: the same fixed supports, 180 eight-step queries and 16 candidates; quality is the only uncertainty.

Check at support tolerance 4/255Queries / 180
Pass Eq. (6)'s common-displacement/orthogonality check0/180
Same correct winner over all seven compatible qualities168/180
Stable and endpoint RMSE changes exceed 1/255142/180

This is finite-domain stability evidence for an actual quantized operation, showing that exact cancellation is not necessary to preserve the winner.

3. AI: using propagation analysis to guide a repair. Following C.3, we localized the four original complete-ranking order errors to the blur approximation before AI. Support-only refitting within the original library repaired all four, keeping candidates and MSE scoring fixed. Among 24 new sources evaluated after model freezing, one reproduced the same error and its repair: the generating-history candidate's endpoint MSE fell from 1.07×10−41.07\times10^{-4} to 1.42×10−161.42\times10^{-16}, while the nearest wrong candidate remained about 9.4×10−69.4\times10^{-6}, correcting the decision.

We will add the calculations, the operation scope of each certificate and the case-level results so that these conditions can be checked against image experiments. We address W2's refinement and correction gains separately in Response 4.

AuthorsOpenReview ↗

Response to nkzW (3/4): Real Platform Workflows and Image Alignment (W1, Q3)

W1 (summary of the workflow concern): Without operations outside the library or real workflows, the C2PA and social-sharing motivation remains unvalidated.

Q3 (summary): Have you tried actual workflows, such as platform recompression? Please also clarify whether pixel-aligned inputs are required.

Thank you for raising these practical questions. We have added an evaluation on real platform processing chains. The original pixel-based version does require pixel alignment, and we have added an inference extension that handles small translations during scoring.

Real platform workflows. We used public V-SMUD chains produced by actual Facebook, Twitter and Flickr upload/download workflows, collected historically. We fixed 52 query sources from 26 cameras absent from the support set, evaluating nine two-step and 27 three-step chains per source. The method, models and split were frozen before acquiring the test queries.

These workflows involve both recompression and resizing. We therefore added native-resolution and processing-geometry adaptation, changing the support representation and candidate replay. Table R3a retains the original pixel version and a control using the same dimension rules. The primary native extension raises two- and three-step recovery to 47.44% and 16.74%.

Table R3a. Historical V-SMUD, 52 sources from 26 support-unseen cameras: 468 two-step and 1,404 three-step queries. Complete-history Top-1 (%). Native adaptation changes support representation, replay and geometry together.

SettingNative adaptationDimension filterTwo stepsThree steps
Original pixels (P0)OffOff26.076.20
Pixels + dimensions (P1)OffOn32.268.83
Native extension (P2, primary)OnOn47.4416.74
Native, no filter (secondary)OnOff56.6221.94

Native representation and geometry adaptation improve recovery even under the same dimension rules. These are results for a method extension. The no-filter row is a prespecified secondary control. Three-step recovery remains difficult, and the hard dimension rules excluded 24.93% of the true three-step histories. The dimension rules need to account for the sizes produced by actual platform workflows.

Pixel alignment. The original comparison is sensitive to misalignment: shifting only the observed endpoint horizontally by one or two pixels reduced recovery on 180 conventional eight-step queries from 100% to 62.22% and 19.44%. Our extension searches the same translation grid (±3 pixels, 0.5-pixel steps) for each candidate prediction, aligning it to the observation before comparing history scores. Models and supports stay fixed; true histories and applied shifts are not used.

After freezing the rule, we tested 48 new images, 24 each from Flickr8k and COCO. Table R3b shows recovery rising from 252/528 to 528/528 under the primary shifts. Input-pair MSE alignment estimates one shared displacement from the two input images.

Table R3b. ImageMagick eight-step queries, 16 candidates, identical fitted operators/supports and the same 48 images across methods. Shift counts are repeated conditions on those images.

Alignment methodAligned (48)11 shifts (48×11)Two off-grid shifts (48×2)
No registration48/48252/52857/96
Input-pair MSE47/48515/52893/96
Candidate predictions48/48528/52895/96

The extension preserved every originally aligned decision and recovered every query under the primary test shifts; one off-grid case remained incorrect. Thus, the original version's pixel-correspondence requirement should be explicit. For the tested small translations, the extension handles misalignment during inference without requiring the true displacement.

These experiments test history recovery after actual social sharing and the effect of small misalignment on decisions. We will state the respective conditions for platform-geometry adaptation and the translation extension, and discuss image-based inference as a complement to partial C2PA provenance records.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Anonymous Code and Reproduction Materials

We thank the reviewers and the Area Chair (AC) for their careful reading.

Your comments prompted us to further examine and clarify ChainPrint’s ability to recover image processing histories from a few anonymous examples, and its practical value. We have presented the additional experiments and analyses in our individual responses.

We also recognized the need to share our implementation so that you can run the method and check the key results directly. We have therefore updated the supplementary material with anonymous code, fitted parameters, and reproduction instructions. You can inspect what information the estimator receives and how history decisions are made.

We hope you will consider the updated supplementary material alongside our responses. We would be happy to discuss any remaining questions.

AuthorsOpenReview ↗

Revised manuscript uploaded

Dear Area Chair and Reviewers,

We have uploaded the revised manuscript. In the PDF, blue text marks revisions. Newly added passages and sections also have a blue vertical bar beside them.

Thank you for your time and feedback.

AuthorsOpenReview ↗

Discussion Summary for the Area Chair and Reviewers

Dear Area Chair and Reviewers,

As the rebuttal period closes, we would like to leave one concise synthesis for the discussion phase.

ChainPrint proposes and studies a new problem: history-decision recovery from a small number of anonymous examples. Given a few before/after pairs, two image versions, and candidate processing histories, the goal is to determine which history best explains the observed change before every individual operation has been fully identified.

The central finding is that history decisions may require less information than complete operation identification. Example-compatible operation models can support a stable history decision even when the underlying operation model is not unique.

The rebuttal tested this idea under the main conditions raised in review.

  • With the function library fixed but the JPEG implementation changed, refreshing anonymous examples improved recovery from 128/180 to 172/180.
  • For an SD1.4 editor absent from the original library, substitutes fitted from anonymous examples improved recovery from 5/24 to 14/24 without supplying the editor executable.
  • We also added computable stability tests, finite-domain JPEG checks, error-propagation analysis and repair, historical platform-chain evaluation, image-alignment experiments, and a follow-up editor test.

The revised manuscript distinguishes fixed-library adaptation, learned substitutes, and adding a matching editor as an explicit candidate. It reports same-prior Joint fitting, preparation cost, candidate availability, and platform conditions separately.

The resulting contribution is a decision-level framework for recovering processing histories from limited operation evidence under an explicit candidate space. The revised manuscript, complete responses, anonymous code, fitted parameters, and reproduction materials provide the full record for the discussion.

Best regards,
The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.