{"id":"1f005dad-1a70-4f9e-8aeb-4c9451353cab","arxiv_id":"2507.22601","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A video deepfake detector for eKYC that combines temporal identity-vector differences with differences against a registered photo, and shows robustness to image degradation.","lead":"This paper builds a deepfake detector for selfie-style identity verification that compares identity features in a video against a trusted registration photo, and tracks how those features change over time. It reports better robustness to blur, noise, and compression than several published baselines, at the cost of lower accuracy on clean cross-dataset video.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The eKYC deployment claim hinges on an untested auxiliary-image assumption: the registered ID photo is never represented in experiments, since auxiliary images are always same-domain video frames.","rationale":"The reader's weakest assumption and my leading concern coincide: the method's eKYC motivation depends on using a genuine registered ID document, but the experiments substitute same-domain video frames as auxiliary images. This is not an internal inconsistency but a deployment-premise gap. It is load-bearing because ADC is one of the three named contributions and the paper's stated application is eKYC; if the auxiliary image is not document-like, the reported AUCs may not transfer. I also noted the abstract's 21.5 pp cross-dataset gain overstates the largest table value (19.2 pp on DFDCp) and the lack of error bars, but these are secondary. The paper has genuine strengths: public code, fixed identity-disjoint splits, and a clear demonstration that larger-scale face recognition improves both performance and degradation robustness. My proposed test is a concrete way to settle whether the deployment premise survives; without it, the claimed eKYC robustness remains conditional.","tokens_in":13881,"tokens_out":4423,"duration_ms":52212,"concrete_test":"Use the released code to evaluate a protocol where auxiliary images are transformed to mimic ID documents: apply perspective/roll-pitch-yaw variation, print-and-scan or high JPEG artifacts, low resolution (e.g., ~56 px face), and lighting shifts to the same-identity auxiliary frames, or use the eKYC-DF registered-ID images if available. Measure video-level AUC and false-positive rate on real videos from KoDF/CDF/DFDCp and compare against the current same-domain auxiliary results. If AUC drops by more than ~5 points or FPR exceeds an eKYC-acceptable threshold, the deployment claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires the Auxiliary Difference Comparison (ADC, Eq. 1) to distinguish genuine users from attackers. In the intended deployment, the auxiliary is an ID-document photo. Yet Section IV-A1 explicitly says that 'the input video is captured in real-time during authentication, whereas the auxiliary image is a photo printed on an identity document' and that the input video and auxiliary image 'often differ visually.' Despite this, every experiment samples the auxiliary image from a different video of the same identity (KoDF, CDF, DFD, DFDCp), so pose, lighting, resolution, and capture noise are far more similar to the input video than a printed ID photo would be. Under the actual eKYC premise, a genuine user's f(video) - f(ID photo) differences will be large and noisy; the trained RNN has never seen such inputs, so it may misclassify genuine users as fake or narrow the real/fake margin. Robustness to input-video degradation (Figs. 3-6, Table I) does not address robustness to auxiliary-video mismatch. The eKYC-specific conclusion is therefore not validated by the reported experiments.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a video-level deepfake detector for eKYC that combines Temporal Difference Comparison (TDC) of identity vectors across frames with Auxiliary Difference Comparison (ADC) against a registered image, feeding both into an RNN with a classification head; the feature extractor is a pretrained face recognition model, and the authors evaluate the effect of extractor scale. Experiments are performed on KoDF for training and in-dataset testing, with cross-dataset evaluation on Celeb-DF v2, DFD, and DFDCp, plus robustness tests under six image degradations. The central claimed contributions are comprehensive detection of face swapping and face reenactment, improved accuracy from the registered image, and robustness to unseen degradation, with source code publicly released.","tokens_in":14066,"tokens_out":4632,"duration_ms":54394,"significance":"If the claims hold, the paper makes a practical contribution to eKYC deepfake defense by showing that an enrollment image can be leveraged for identity-consistency detection, and by demonstrating a positive correlation between face-recognition feature quality and deepfake-detection robustness. The systematic robustness evaluation across six corruptions, the feature-extractor ablation, and the release of source code are genuine strengths. However, the eKYC-specific conclusion depends on an auxiliary-image assumption that the experiments do not actually test, and the headline quantitative claims contain an inconsistency that must be resolved.","major_comments":[{"comment":"The eKYC premise stated in Section IV-A1 is that the input video is captured in real-time while the auxiliary image is a photo printed on an identity document, and that these 'often differ visually.' Yet in every experiment the auxiliary image is sampled from a different video of the same identity within the same dataset, so the ADC branch is never exposed to the document-photo-to-video domain gap that motivates the method. Robustness experiments corrupt only the input video, not the auxiliary image. To support the eKYC claim, please report experiments where the auxiliary image comes from a genuinely different capture domain (e.g., document scans, lower resolution, different lighting/pose), or at least analyze the distribution of ADC identity-vector distances for genuine users under such mismatch; without this, the central deployment claim is not validated.","section":"Section IV-A1 and Eq. (1)"},{"comment":"The introduction states that the method outperforms TI2Net by 1.5 and 21.5 percentage points in in-dataset and cross-dataset evaluation, respectively. Table III, however, shows per-dataset gaps of 11.72 points on Celeb-DF v2, 11.66 points on DFD, and 19.21 points on DFDCp, and the average gap is about 14.2 points. The 21.5-point figure is not reproducible from the reported tables; please correct the statement or specify the exact computation used.","section":"Section I and Table III"},{"comment":"The face feature extractor is pretrained on WebFace12M, a large web-scraped dataset, and the paper does not address the possibility that WebFace12M contains identities that overlap with KoDF, Celeb-DF v2, DFD, or DFDCp evaluation subjects. Because the method's representation is explicitly identity-based, unseen-identity generalization is central to the cross-dataset claim. Please report whether overlap checks were performed, or temper the generalization claim accordingly.","section":"Section IV-A3 and Section IV-D"},{"comment":"The headline in-dataset comparison is based on a single random split of KoDF and one random video/feature sampling, with no error bars or repeated runs. The reported 'all' AUC of 95.70 is only 0.3 points above SBI (95.40) and 1.5 points above TI2Net (94.20), so without variance estimates it is not clear that the proposed method's advantage is statistically meaningful. Please provide multiple seeds with mean and standard deviation, or otherwise report per-run values.","section":"Section IV-A3 and Table II"}],"minor_comments":[{"comment":"The column header 'DFC' should be 'DFD' to match the dataset name used elsewhere in the paper.","section":"Table VII"},{"comment":"The text says 'We randomly sample 403 real videos and 850 fake videos from the KoDF dataset,' but KoDF contains tens of thousands of videos; please clarify whether this sampling is done before or after the train/validation/test split and whether the same videos are reused for validation and test.","section":"Section IV-A3"},{"comment":"The robustness figures are difficult to read at the plotted scale; consider providing a supplementary table with the numeric AUC values for each method, corruption type, and severity level.","section":"Figures 3-6"},{"comment":"References [10], [11], and [13] appear to refer to the same eKYC deepfake papers with overlapping titles; please consolidate or disambiguate them.","section":"References"},{"comment":"The statement that 'ADC alone surpasses competing methods on four of six datasets' is not directly supported by Table V, which only compares embedding types and does not include competing methods; please clarify which table and which comparison is intended.","section":"Section V"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a security/computer-vision venue and the code release is a plus. The main issue is not the core method but the gap between the eKYC deployment claim and the auxiliary-image proxy used in experiments; this is addressable with additional experiments or a clearly scoped claim. The numerical inconsistency in the abstract's 21.5-point claim should also be fixed. I would not reject the paper, but I would require the domain-mismatch evaluation before accepting the eKYC-specific conclusion."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a legitimate empirical paper on identity-based deepfake detection, and the feature-extractor scaling result is genuinely useful; but the eKYC deployment claim rests on an auxiliary-image assumption the experiments never actually test. The paper is better than the abstract suggests in some ways and worse in others.\n\nWhat's new: combining temporal identity differences (TDC) with differences against a registered image (ADC) plus a large-scale AdaFace encoder. The ablations in Table V are clean: ADC alone catches reenactment, TDC catches swapping, the combination handles both. The scaling study (Table IV, Figs. 5-6) is the strongest part—showing that a better face recognition model improves both AUC and robustness against degradation, which is a useful empirical result for the community. Code is public and the KoDF split is released, so the results are reproducible.\n\nNow the soft spots. The biggest one: the auxiliary image in every experiment is a frame from another video of the same identity, not an ID-document photo. Section IV-A1 explicitly says the eKYC input video and an ID photo 'often differ visually,' yet no experiment simulates that mismatch. The ADC input distribution in the field will include print artifacts, lighting and pose gaps, and the RNN has never seen such inputs. So the central eKYC-specific conclusion is not validated by the reported experiments. That doesn't sink the method as a general detector, but it means the title overreaches.\n\nSecond, the abstract's '21.5% points' cross-dataset improvement over TI2Net doesn't match Table III; the largest gap there is 19.2pp. The 21.5pp appears in Table I, a Gaussian-blur robustness table at pristine distortion, which is not the same claim. This should be corrected.\n\nThird, the statement that ADC is 'not previously addressed' is contradicted by the paper's own references [11] and [13], which are eKYC deepfake detection papers with face registration. Incidentally, [11] duplicates [10] and [13] duplicates [12]; the citation list needs a cleanup.\n\nFourth, all AUCs come from one split with no error bars. With a public split that's acceptable for a first report, but the differences among methods are often a few points, so error bars would materially raise confidence.\n\nBottom line: this is a solid, reproducible contribution to identity-based deepfake detection, worth citing for the encoder-scaling result and the ADC idea. A serious referee should engage with it, but the authors need to fix the abstract, acknowledge the auxiliary-domain gap, and ideally add a simulation of ID-photo mismatch. My own verdict would be major revision, not rejection.","headline":"Useful encoder-scaling results and a clean ablation, but the eKYC deployment claim is untested because the auxiliary image is never an ID photo.","tokens_in":14688,"tokens_out":4010,"would_cite":true,"duration_ms":41893,"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":"The paper claims that a deepfake detector for eKYC systems can catch both face-swapping and face-reenactment attacks by comparing face-recognition identity vectors across video frames and against a registered photo, while staying accurate…","keywords":["deepfake detection","eKYC","face swapping","face reenactment","identity vectors","registered image","temporal inconsistency","image degradation robustness"],"falsifier":"Give the trained detector a real-user video paired with a genuine ID document photo taken under office lighting, with print texture and a different camera, and measure the video-level AUC on real users: if the ADC stream pushes genuine users' scores closer to the fake distribution and AUC collapses, the registered-image contribution to eKYC robustness is not as claimed.","tokens_in":13612,"feed_emoji":"🎭","tokens_out":8526,"duration_ms":85016,"temperature":0.7,"pith_summary":"The paper claims that a deepfake detector built for electronic Know Your Customer (eKYC) systems can flag both face-swapping and face-reenactment attacks by checking identity in two ways at once: how much a person's face-recognition identity vector jumps between consecutive video frames, and how far those frame vectors sit from a registered reference photo of the same claimed person. The authors show that the second check, which they call Auxiliary Difference Comparison, is what gives the method an edge on reenactment attacks, while the first check handles swapping, and that the two together beat the temporal-identity baseline on both in-dataset and cross-dataset tests. They also report that swapping the face encoder for one trained with AdaFace loss on the WebFace12M dataset raises in-dataset AUC from roughly 81% to 95% and sharply reduces accuracy loss under blur, noise, and compression. If the claim holds, eKYC systems already possess the extra signal they need—the enrolled identity photo—and the main remaining design choice is which face recognition backbone to use.","feed_headline":"Registered-photo check catches face-swap and reenactment deepfakes","feed_subtitle":"Temporal identity-vector differences, plus a check against a trusted photo, keep eKYC detection accurate on degraded video.","key_machinery":"The carrying object is the identity vector: a fixed-dimension embedding of a face produced by a face recognition network. The detector feeds a bidirectional GRU with two concatenated difference sequences—temporal differences between consecutive frames and frame-by-frame differences against a registered reference image—and the GRU's final hidden state is classified by two fully connected layers. The second carrying element is the feature extractor itself: a ResNet100 trained with AdaFace loss on WebFace12M, whose identity vectors remain stable when input frames are blurred, noisy, compressed, or altered in saturation and contrast.","core_discovery":"On the paper's own terms, the central discovery is that identity-vector dynamics, when measured as differences rather than absolute embeddings, are a sufficient and unusually degradation-tolerant basis for eKYC deepfake detection. The method computes two parallel difference sequences from a face recognition extractor: temporal differences between successive frames (TDC) and per-frame differences against a registered genuine image (ADC), concatenates them, and classifies the sequence with a bidirectional GRU plus a two-layer head. In KoDF in-dataset evaluation the full model reaches 95.70% video-level AUC across all fake types, ahead of both the TDC-only TI2Net baseline and the frame-level SBI baseline, and the paper reports a 21.5-point advantage over TI2Net in cross-dataset evaluation. The paper additionally demonstrates, by comparing three face encoders, that the accuracy and corruption robustness of the detector track the scale and quality of the face recognition training data, with the AdaFace/WebFace12M encoder giving both the best detection and the smallest AUC declines under six types of image degradation.","pith_inferences":["A testable extension would swap the auxiliary image source from a same-identity video frame to an actual ID document scan or photo; if the capture-condition gap is large, ADC may need a calibration step or a learned tolerance before it helps in deployed eKYC.","The success of ADC suggests a broader design principle: any system that has a trusted reference image of the claimed identity can use identity-vector distance to that reference as an attack signal, which could apply to liveness checks, speaker verification, or document-holder matching.","Because the paper shows a scaling trend from ArcFace/MS1MV2 to AdaFace/MS1MV2 to AdaFace/WebFace12M, a natural next experiment is to measure whether even larger face recognition backbones continue to raise both detection AUC and degradation robustness, and whether gains saturate.","The random-sampling result hints that a simpler, order-agnostic aggregation of identity differences could match the GRU's accuracy at lower latency, which would be worth testing for real-time eKYC liveness."],"forward_implications":["eKYC systems can use the enrolled identity photo as an auxiliary signal without adding a new capture step, and the ADC branch gives the detector a specific advantage on face-reenactment attacks.","Because detection accuracy and robustness tracked the face encoder's training scale in the paper's comparisons, the method will likely continue to improve as face recognition models are trained on larger datasets.","The small AUC declines under six degradation types imply identity-vector-based detection is suitable for real eKYC video, where blur, noise, compression, and block distortion are common.","Cross-dataset results on Celeb-DF v2, DFD, and DFDCp indicate the detector generalizes to manipulation methods and identities not seen in training, and its DFDCp result shows particular strength on post-processed, low-quality videos."],"supporting_citations":[{"why":"Supplies the temporal identity inconsistency baseline and the TDC formulation that the proposed method extends with the auxiliary difference branch.","marker":"[32]"},{"why":"Provides the Korean DeepFake Detection Dataset, whose multiple videos per identity make auxiliary image sampling possible for training and evaluation.","marker":"[27]"},{"why":"Provides the AdaFace loss used to train the face feature extractor that gives the detector both accuracy and corruption robustness.","marker":"[24]"},{"why":"Provides the WebFace12M training data for the large-scale face recognition backbone used in the main experiments.","marker":"[55]"},{"why":"Serves as the primary cross-dataset evaluation benchmark (Celeb-DF v2) used to measure generalization and degradation robustness.","marker":"[31]"},{"why":"Frame-level self-blended-images baseline that the method is compared against for detection accuracy and robustness to degradation.","marker":"[41]"},{"why":"Video-level lip-movement baseline used to demonstrate the robustness advantage under Gaussian blur and other degradations.","marker":"[18]"}],"fun_headline_variants":["Identity-vector differences catch swap and reenactment deepfakes","Registered-photo check boosts deepfake detection in eKYC","Temporal and registered identity checks defeat deepfakes","eKYC deepfake detector stays accurate under video degradation","Trusted photo plus temporal identity check spots deepfakes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The registered image really is the same person as the video and is similar enough in pose, lighting, and capture conditions that a genuine user's identity-vector differences stay close to zero; the paper only tests this with frames from other videos of the same identity, not with actual ID document photographs.","fun_headline_variants_meta":{"raw":{"variants":["Identity-vector differences catch swap and reenactment deepfakes","Registered-photo check boosts deepfake detection in eKYC","Temporal and registered identity checks defeat deepfakes","eKYC deepfake detector stays accurate under video degradation","Trusted photo plus temporal identity check spots deepfakes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000751,"raw_usage":{"total_tokens":3361,"prompt_tokens":984,"completion_tokens":2377,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":600,"completion_tokens_details":{"reasoning_tokens":2293}},"tokens_in":600,"tokens_out":2377,"duration_ms":19966,"temperature":1.0,"reasoning_tokens":2293,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T11:28:49.386925+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Give the trained detector a real-user video paired with a genuine ID document photo taken under office lighting, with print texture and a different camera, and measure the video-level AUC on real users: if the ADC stream pushes genuine users' scores closer to the fake distribution and AUC collapses, the registered-image contribution to eKYC robustness is not as claimed.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the temporal identity inconsistency baseline and the TDC formulation that the proposed method extends with the auxiliary difference branch."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Korean DeepFake Detection Dataset, whose multiple videos per identity make auxiliary image sampling possible for training and evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the AdaFace loss used to train the face feature extractor that gives the detector both accuracy and corruption robustness."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the WebFace12M training data for the large-scale face recognition backbone used in the main experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Serves as the primary cross-dataset evaluation benchmark (Celeb-DF v2) used to measure generalization and degradation robustness."},{"cited_title":"Shiohara and T","cited_arxiv_id":null,"evidence_quote":"Frame-level self-blended-images baseline that the method is compared against for detection accuracy and robustness to degradation."},{"cited_title":"Haliassos, K","cited_arxiv_id":null,"evidence_quote":"Video-level lip-movement baseline used to demonstrate the robustness advantage under Gaussian blur and other degradations."}],"review_version":1}