{"id":"9ee0ad1c-9085-4d2b-b720-965e109404a7","arxiv_id":"2501.16227","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"PDC-ViT feeds pixel-difference convolution features into a Vision Transformer and reports 83 to 94 percent accuracy on five public source-camera identification datasets.","lead":"A new forensic method combines two existing ideas, pixel-difference convolutions and a Vision Transformer, to identify which smartphone camera took an image or video frame. The authors report the highest accuracy on several public source-camera datasets, though the training and test setup and the missing code make the numbers hard to verify.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported accuracies rely on a random frame-level split and uncontrolled baselines; without a video-disjoint test partition, the headline numbers may reflect video-content memorization rather than true source-camera generalization.","rationale":"The reader's conditional verdict is appropriate. I agree with its weakest-assumption choice: split leakage is the point most capable of invalidating the headline. The paper's own text (§4.1, §4.3) never rules out same-video frames in train and test; for video forensics this is a known source of overoptimistic accuracy. My addition is that even the cleanest split would not by itself validate \"outperforms SOTA\", because Tables 2-5 mix device counts, resolutions, and protocols; a controlled head-to-head rerun of at least one strong baseline is needed. The internal numerical contradictions make independent reproduction necessary. These are fixable issues: release code and a precise split, rerun baselines, report error bars. Hence no verdict change; conditional remains.","tokens_in":16565,"tokens_out":5394,"duration_ms":51564,"concrete_test":"Re-run the Vision and Video-ACID experiments with a video-disjoint split: for each device, assign complete videos (not frames) to train/validation/test, e.g., 80% of videos per device for training/validation and 20% for testing, and ensure no frames from a held-out video appear in training. Repeat with 5 random seeds and compare top-1 accuracy against the reported random-frame split. If the video-disjoint accuracy drops materially (say >3 percentage points on Vision or >5 on Video-ACID), frame-level leakage is confirmed and the headline accuracies are inflated. As a secondary check, independently compute the Daxing 224x224 number to decide whether it is 84.06 or 82.81, and re-run one baseline (e.g., DHDN [21] at 512x512) under the same split protocol to verify the claimed margin.","verdict_should_be":"UNCHANGED","load_bearing_attack":"To support \"outperforms SOTA\", PDC-ViT must be shown to generalize to new captures from known cameras, not to memorize content of particular videos. The protocol in §4.1 and §4.3 is an 80/20 random split with images \"selected randomly from each set of cameras\", and the paper never states that all frames of a video are kept in one partition. For Vision, Daxing, Socrates, QUFVD, and Video-ACID, frames extracted from the same recording share scene content, lighting, compression, and temporal noise; if any of those frames appear in both training and test, a classifier can separate cameras by video-specific cues and inflate accuracy. The comparisons in Tables 2-5 deepen the problem: baseline numbers come from different device subsets (e.g., 11 vs 28 vs 35 devices on Vision), different image resolutions, and different train/test protocols, so the claimed margins are not established under a controlled protocol. Internal inconsistencies (Daxing: \"82.81\" in §4.4 vs \"84.06\" in Table 3; APDC/RPDC accuracies reversed in §4.6) further mean the numbers cannot be trusted until reproduced. Thus the central claim is currently unverified, not refuted; a clean video-disjoint evaluation could settle it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PDC-ViT, a source camera identification method that computes Angular and Radial Pixel Difference Convolution features from input images and feeds them to a Vision Transformer for classification. The authors report closed-set identification accuracies of 94.30% on Vision, 84.06% on Daxing, 94.22% on Socrates, 92.29% on QUFVD, and 74.77% on Video-ACID, and they compare these numbers with previously published results on the same datasets to claim state-of-the-art performance. The manuscript also provides confusion matrices and t-SNE visualizations for some datasets.","tokens_in":16820,"tokens_out":5522,"duration_ms":43869,"significance":"If the reported results survive a clean evaluation, the idea of using PDC as a feature extractor for ViT-based source camera identification is a reasonable engineering contribution that avoids explicit PRNU estimation and could be useful for forensic applications. The paper is honest about its limitations (Section 5) and uses public datasets, which makes the claims checkable in principle. However, the current validation does not yet establish the central 'superiority' claim: the split protocol may allow frame-level leakage, the baselines are compared under heterogeneous protocols, and several numbers are internally inconsistent.","major_comments":[{"comment":"The evaluation does not confirm that the training/validation/test split is disjoint at the video or device level. Section 4.1 states only that 'we used 80% of each dataset for Training/validation and 20% for testing,' and Section 4.3 says 'the images are selected randomly from each set of cameras'; for video-based datasets such as Vision, Daxing, QUFVD, and Video-ACID, this procedure can place frames from the same video or recording session in both training and test, allowing the model to exploit video-specific content and inflate accuracy. A video-disjoint (or at least session-disjoint) split is necessary to support the claim that PDC-ViT generalizes to new captures from known cameras; please describe the exact partition rule and rerun the experiments under a disjoint partition.","section":"§4.1, §4.3"},{"comment":"The comparisons against state-of-the-art methods are not controlled. In Table 2, baseline accuracies are cited for different numbers of devices (11, 28, or 35), different input resolutions (32x32 to 480x800), and different training protocols; Table 4 gives no protocol information for the baselines; and Table 5 mixes resolutions and does not specify the device subset used for evaluation. Since accuracy in this task depends strongly on class count, image size, and split, the claimed margins over prior work are not established by the presented numbers. Please either re-run the baselines under identical conditions or clearly justify why cross-paper comparisons with differing protocols are valid.","section":"Tables 2–5"},{"comment":"Several reported numbers are internally inconsistent. Section 4.4 reports a Daxing accuracy of 82.81 but Table 3 reports 84.06 for the same setting; Section 4.6 says PDC-ViT, RPDC-ViT, and APDC-ViT achieved 92.29%, 85.47%, and 83.36%, respectively, while Table 5 lists RPDC-ViT as 83.36 and APDC-ViT as 85.47, reversing the two variants. These contradictions need to be resolved and the correct table/text values verified before the accuracy claims can be assessed.","section":"§4.4, §4.6, Table 3, Table 5"},{"comment":"All reported accuracies are single-run numbers without error bars or repeated trials, and the stability claim in Section 4.7 is based on one run. Given that some margins over the second-best method are small (e.g., 1.91% in Table 4), statistical significance is unknown. Please provide mean and standard deviation over multiple training seeds and, for video datasets, per-video aggregated metrics to show that the improvement is not due to a particular train/test draw.","section":"§4.4–§4.7"}],"minor_comments":[{"comment":"In the sentence 'the use of pixel difference convolution exploding the angular and radial differences ... using 3 × 3 kernels for Angular PDC (APDC) and 5 × 5 kernels for Radial PDC (APDC)', the second occurrence of '(APDC)' should be '(RPDC)'.","section":"§3.1"},{"comment":"The sentence 'In the case of the iPhone 6S (Plus), 23 different smartphone models are available' is unclear; it seems to describe the Daxing dataset but the wording should be revised.","section":"§4.2"},{"comment":"The text uses 'FNTs' where 'FNRs' is intended, and the abbreviation 'Video W A PRNU' is never defined; please define it at first use (likely 'weighted averaging PRNU' from reference [27]).","section":"§4.7"},{"comment":"The caption of Figure 4 lists 'Vision dataset' in its title while the subfigures and text refer to QUFVD; please correct the caption to match the content.","section":"Figure 4"},{"comment":"The 'Overall' row in Table 6 does not specify whether the FNR and FPR are macro-averaged over devices or micro-averaged over all test samples; please clarify the averaging procedure.","section":"Table 6"},{"comment":"The code availability statement says code 'will be published when the paper is accepted'; since the evaluation protocol is a central issue, releasing the exact data split would substantially strengthen reproducibility.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":"The paper is a standard closed-set camera identification paper whose central novelty is the PDC+ViT combination. The main risk is that the benchmark numbers may be inflated by an insufficiently strict split; if that is confirmed, the SOTA claim collapses. I would recommend that the editor request the exact split code or a detailed video/session-disjoint protocol, and possibly ask an independent reviewer with multimedia forensics expertise to assess the comparison tables. The paper's internal inconsistencies also suggest the authors should carefully re-check their recorded results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the quick take: this paper combines Pixel Difference Convolution (angular and radial) with a Vision Transformer for closed-set source camera identification. That's a genuine application-level novelty—no one else feeds PDC features into a ViT for this task—and the method is plausible in principle: PDC emphasizes local pixel differences that carry sensor artifacts, and ViT can capture long-range dependencies. The paper also evaluates on five datasets and compares against a wide range of baselines, with confusion matrices and t-SNE plots. That is real work.\n\nBut the evidence is not as solid as the claims. The central problem is the split protocol. Section 4.3 says test images are 'selected randomly from each set of cameras.' That does not tell us whether frames from the same video can appear in both training and test. For Vision, Daxing, QUFVD, and Video-ACID, frames extracted from a shared recording share scene content, lighting, compression, and temporal noise. If any video spans both partitions, the classifier can separate cameras by video-specific cues, and accuracy is inflated. The paper simply never confirms a video-disjoint or camera-disjoint split, and the stress-test note is right that this is load-bearing. The comparisons are also heterogeneous: Table 2 mixes results from 11, 28, and 35 devices at different resolutions, so the claimed margins over prior work are not established under a controlled protocol.\n\nThere are also smaller but telling inconsistencies. The Daxing accuracy appears as 82.81 in Section 4.4 and 84.06 in Table 3. In Section 4.6 the text says RPDC-ViT scores 85.47 and APDC-ViT scores 83.36, but Table 5 lists them the other way around. No error bars or repeated runs are reported, and the code is only promised 'when the paper is accepted.'\n\nNone of this is fatal to the idea. Pixel-difference features are a reasonable choice for this task, and the pipeline is simple and easy to reproduce once the split is pinned down. But right now the headline numbers are unverified, not refuted. A clean video-disjoint evaluation with controlled baselines and variance estimates could settle it. I would not cite this in its current form, and I would be cautious about building on the numbers. It does deserve a serious referee, though: the topic is relevant, the combination is novel enough, and the flaws are fixable.","headline":"Plausible PDC+ViT combination for source camera ID, but the evaluation is under-specified and the headline numbers are not yet trustworthy.","tokens_in":17348,"tokens_out":3071,"would_cite":false,"duration_ms":26984,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"PDC-ViT claims that pixel-difference convolution features fed into a Vision Transformer outperform prior source camera identification methods on five public benchmarks.","keywords":["source camera identification","Pixel Difference Convolution","Vision Transformer","angular pixel difference","radial pixel difference","multimedia forensics","camera model identification","deep learning"],"falsifier":"Re-run the evaluation with a strict camera-disjoint or video-disjoint split, training on some cameras or videos and testing only on cameras or videos never seen in training, and compare the accuracy. If PDC-ViT's margin over baselines collapses or drops substantially, the claimed generalization to source cameras is an artifact of same-video frame overlap.","tokens_in":16388,"feed_emoji":"📷","tokens_out":9630,"duration_ms":71401,"temperature":0.7,"pith_summary":"The paper proposes PDC-ViT, a source camera identification method that replaces content-based feature extraction with Pixel Difference Convolution (PDC), computing angular and radial differences between neighboring pixels, and feeds those features into a Vision Transformer for classification. The claim is that this combination captures sensor-specific pixel-level variations better than conventional CNN backbones, and the reported results support that: 94.30% on Vision, 84.06% on Daxing, 94.22% on Socrates, 92.29% on QUFVD, and 74.77% on Video-ACID, each beating previously published methods on those benchmarks. A sympathetic reader would care because source camera identification is used in digital forensics to trace images and videos to the device that captured them, and the paper argues that pixel-difference features plus attention give a content-independent, high-accuracy route to that trace.","feed_headline":"Pixel-difference features push camera ID to 94%","feed_subtitle":"Angular and radial pixel-difference features feed a Vision Transformer that beats prior methods on five datasets.","key_machinery":"The load-bearing mechanism is Pixel Difference Convolution, which replaces the vanilla convolution sum with weighted differences between paired neighboring pixels: $Y = \\sum_{(x_i,x_i') \\in P} w_i \\,(x_i - x_i')$. Angular PDC uses 3x3 kernels that follow a circular ordering of pixel pairs; Radial PDC uses 5x5 kernels that pair a center pixel with radially arranged neighbors. These PDC feature maps are concatenated and embedded as 64x64 patches into a Vision Transformer with six encoder layers, sixteen attention heads, and embedding dimension 1024. The PDC stage is what suppresses scene content and exposes sensor-specific micro-patterns; the ViT stage is what classifies those patterns globally. The paper's comparison of APDC-only, RPDC-only, and combined variants isolates the contribution of each difference geometry.","core_discovery":"On its own terms, the paper's central claim is that source camera identity is best learned from pixel-difference statistics rather than from scene content. The proposed PDC-ViT computes Angular PDC with 3x3 kernels and Radial PDC with 5x5 kernels, concatenates their feature maps, and feeds the resulting PDC features as patches to a Vision Transformer. Across five public datasets, the authors report accuracy of 94.30% (Vision), 84.06% (Daxing), 94.22% (Socrates), 92.29% (QUFVD), and 74.77% (Video-ACID with added devices), exceeding the best prior results they compare against on each benchmark. They also report that combining angular and radial branches outperforms either branch alone, with 92.29% versus 85.47% and 83.36% on QUFVD, and that PDC feature clusters are more separated in t-SNE visualization than ResNet101 features.","pith_inferences":["A camera-disjoint evaluation is the natural next test: if the 94.30% Vision accuracy persists when training and test sets share no camera, the method would generalize to unseen devices of the same models; the paper's random 80/20 split does not establish this.","Because PDC suppresses scene content, the same backbone might transfer to manipulated or recompressed images more gracefully than content-based CNNs, but the paper does not test compression or anti-forensic manipulation.","The architecture suggests a cheap ablation probe: keeping the same ViT, replacing PDC with ordinary convolution at matched parameter count would isolate whether the pixel-difference operation itself, rather than the larger effective receptive field, drives the gain.","The reported t-SNE separation of PDC features against ResNet101 features hints that an open-set or verification variant of PDC-ViT could work, but the paper only evaluates closed-set classification."],"forward_implications":["If the reported numbers hold, pixel-difference features plus a Vision Transformer are enough to identify source cameras at state-of-the-art accuracy on closed-set benchmarks with up to 35 devices.","Combining angular and radial PDC branches improves over either branch alone, so the two difference geometries carry complementary identity information.","The method reaches competitive accuracy on both image datasets and video-frame datasets, suggesting the pixel-difference signal survives video compression and scene variation.","On Video-ACID, the overall false negative rate of 23.4% and false positive rate of 0.92% are lower than the PRNU-based baselines compared against, which would mean fewer missed identifications and fewer false attributions in forensic use."],"supporting_citations":[{"why":"supplies the random pixel difference convolution formulation that the PDC backbone extends to angular and radial variants.","marker":"[29]"},{"why":"supplies the angular, radial, and central pixel difference kernel designs used in the feature extraction blocks.","marker":"[30]"},{"why":"supplies the Vision Transformer encoder architecture that classifies the PDC features.","marker":"[33]"},{"why":"provides the Vision dataset with 35 devices, the main benchmark for the reported 94.30% accuracy.","marker":"[4]"},{"why":"provides the Daxing smartphone dataset used for the 84.06% evaluation.","marker":"[5]"},{"why":"provides the SOCRatES smartphone dataset used for the 94.22% evaluation.","marker":"[6]"},{"why":"provides the QUFVD video dataset used for the 92.29% evaluation.","marker":"[7]"},{"why":"the DHDN method is the strongest baseline the paper compares against on the Daxing dataset.","marker":"[21]"},{"why":"the XceptionNet result is the strongest baseline the paper compares against on the Vision dataset.","marker":"[23]"},{"why":"the forensic similarity result is the second-best baseline the paper compares against on the Socrates dataset.","marker":"[16]"}],"fun_headline_variants":["Pixel-difference features feed ViT for camera ID, 94% accuracy","PDC-ViT uses pixel-difference convolution for 94% camera ID","Camera ID via pixel-difference vision transformer beats baselines","PDC-ViT: pixel-difference features boost camera ID accuracy to 94%","Angular and radial pixel differences boost camera ID to 94%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported accuracy depends on the 80/20 train/test split being free of frame-level leakage: frames from the same video or recording session must not appear in both training and testing, because the model could then recognize a particular video's noise pattern rather than the camera's fingerprint.","fun_headline_variants_meta":{"raw":{"variants":["Pixel-difference features feed ViT for camera ID, 94% accuracy","PDC-ViT uses pixel-difference convolution for 94% camera ID","Camera ID via pixel-difference vision transformer beats baselines","PDC-ViT: pixel-difference features boost camera ID accuracy to 94%","Angular and radial pixel differences boost camera ID to 94%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000972,"raw_usage":{"total_tokens":4190,"prompt_tokens":1059,"completion_tokens":3131,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":675,"completion_tokens_details":{"reasoning_tokens":3032}},"tokens_in":675,"tokens_out":3131,"duration_ms":20693,"temperature":1.0,"reasoning_tokens":3032,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T13:36:13.157388+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the evaluation with a strict camera-disjoint or video-disjoint split, training on some cameras or videos and testing only on cameras or videos never seen in training, and compare the accuracy. If PDC-ViT's margin over baselines collapses or drops substantially, the claimed generalization to source cameras is an artifact of same-video frame overlap.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the random pixel difference convolution formulation that the PDC backbone extends to angular and radial variants."},{"cited_title":"& Liu, L","cited_arxiv_id":null,"evidence_quote":"supplies the angular, radial, and central pixel difference kernel designs used in the feature extraction blocks."},{"cited_title":"”VISION: a video and image dataset for source identification.” EURASIP Journal on Information Security 2017.1 (2017): 1-16","cited_arxiv_id":null,"evidence_quote":"provides the Vision dataset with 35 devices, the main benchmark for the reported 94.30% accuracy."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the Daxing smartphone dataset used for the 84.06% evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the SOCRatES smartphone dataset used for the 94.22% evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the QUFVD video dataset used for the 92.29% evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"the DHDN method is the strongest baseline the paper compares against on the Daxing dataset."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"the XceptionNet result is the strongest baseline the paper compares against on the Vision dataset."},{"cited_title":"Mayer and M","cited_arxiv_id":null,"evidence_quote":"the forensic similarity result is the second-best baseline the paper compares against on the Socrates dataset."}],"review_version":1}