{"id":"b9ad142d-fa4b-428e-9657-2491648ef09e","arxiv_id":"2412.20799","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A deepfake detector that learns to reweight five handcrafted features reports higher cross-dataset AUC than prior baselines, but the mechanism and experiments are not specified enough to verify.","lead":"This paper presents SFE-Net, a deepfake detector that combines five handcrafted image features with a neural network that is supposed to learn which features to trust. It reports better cross-dataset detection than existing baselines, but gives no code or training details, so the improvements cannot be checked.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported cross-dataset margin rests on an unstated SFE-Net evaluation protocol; Section III.E only specifies DeepfakeBench baselines, so Table I's 0.795 vs 0.767 is not a controlled comparison.","rationale":"The reader's weakest assumption and my load-bearing concern are the same: the paper compares SFE-Net against DeepfakeBench baselines without establishing that SFE-Net was evaluated under an identical protocol. Section III.E explicitly attributes baseline numbers to DeepfakeBench weight files, but no corresponding protocol description is given for SFE-Net. This is not a disagreement with a field consensus; it is a missing control in the paper's own comparison. The margin over SPSL (0.795 vs 0.767) is small enough that plausible differences in frame sampling, face cropping, or aggregation could explain it, so the central claim cannot be accepted on the current evidence. The concrete test would settle the concern: reproducing both SFE-Net and SPSL inside DeepfakeBench's harness and checking whether the reported numbers hold. Since the paper provides no code or weights, the test cannot currently be run, which reinforces the REJECT verdict. I do not see a more fundamental internal inconsistency in the architecture as described, but that description is too underspecified to salvage the empirical claim. Therefore the reader's verdict should remain unchanged.","tokens_in":7024,"tokens_out":3420,"duration_ms":35567,"concrete_test":"Run SFE-Net inside the DeepfakeBench evaluation harness: train on the same FF++ c23 split used for [26], apply DeepfakeBench's preprocessing and face alignment, sample frames identically, and compute frame-level AUC with DeepfakeBench's official metric code. As a control, re-run SPSL [20] in the same harness. If the reproduced SFE-Net average differs from Table I's 0.795 by more than ±0.01, or if the reproduced SPSL average differs from 0.767, the reported 0.028 margin is not established. This test requires release of SFE-Net training code and weights, which the paper currently does not provide.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is an empirical superiority claim: SFE-Net achieves 0.795 average frame-level AUC across CDF-v1, CDF-v2, DFD, DFDC, and DFDCP, versus 0.767 for the best baseline SPSL (Table I). For this comparison to be valid, SFE-Net must be evaluated under the same protocol that produced the DeepfakeBench baselines: same FF++ c23 training split, same face alignment/cropping, same frame sampling, same video-level aggregation, and same evaluation code. Section III.E says the baseline numbers come from weight files provided by DeepfakeBench [26], but it never states that SFE-Net was trained or evaluated under that protocol. Section III.A only names the datasets; Section III.C describes the architecture at figure level; Section III.D says the ablation is 'consistent with the evaluation process' without defining that process. Training hyperparameters, optimizer, loss, frame count, and frame-level versus video-level AUC are absent. The reported margin of 0.028 over SPSL is small relative to known protocol sensitivity in cross-dataset deepfake evaluation: face-crop padding, frame selection, and aggregation choices can shift AUC by several points. Therefore the load-bearing condition for the paper's headline claim—that the margin is caused by SFE-Net rather than by a different evaluation pipeline—is unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SFE-Net, a deepfake detector that extracts five handcrafted features (lighting consistency, high-frequency/phase, compression-reconstruction, morphological, and texture cues) and uses a module of LSTM/softmax layers to adaptively weight these features before classification. Experiments trained on FF++ c23 and tested on Celeb-DF-v1, Celeb-DF-v2, DFD, DFDC, and DFDCP report an average frame-level AUC of 0.795, compared with 0.767 for the best baseline (SPSL), and an ablation showing that the SFE module improves a five-feature fusion from 0.771 to 0.795. The manuscript is largely descriptive: the biological analogy is elaborated at length, but the architecture, training protocol, and evaluation details are not specified beyond a figure and two short tables.","tokens_in":7270,"tokens_out":5086,"duration_ms":49654,"significance":"If the reported numbers were obtained under exactly the same protocol as the DeepfakeBench baselines, the result would be a modest but useful empirical contribution: training only on FF++ c23 and evaluating on five held-out datasets is the standard cross-dataset generalization setting, and a 0.028 average-AUC improvement over SPSL is practically interesting. The paper does not provide code, formal derivations, parameter-free constructions, or repeated-run statistics, so the contribution rests entirely on Tables I and II. The biological framing is not itself a technical contribution, and the proposed feature-selection mechanism is plausible but under-specified. The comparison against DeepfakeBench baselines is a good starting point and gives the paper a clear falsifiable empirical claim.","major_comments":[{"comment":"Section III.E / Table I: the headline comparison (SFE-Net 0.795 vs. SPSL 0.767 average frame-level AUC) is not a controlled comparison. The text states only that the baseline numbers use weight files provided by DeepfakeBench [26]; it never states the training split, face cropping/alignment, frame sampling, video-level aggregation, or evaluation code used for SFE-Net. In cross-dataset deepfake evaluation these choices can shift AUC by several points, so the reported margin may be an artifact of protocol rather than of the SFE-Net module. The authors should run SFE-Net inside the DeepfakeBench evaluation harness, or otherwise document the full protocol and demonstrate parity with the baselines.","section":"III.E, Table I"},{"comment":"Section III.D / Table II: the ablation gain of 0.024 (0.771 to 0.795) is reported as a single number with no error bars, no number of runs, and no significance test. Since the variability of each row is not quantified, the claim that SFE-Net offers a consistent improvement over the static five-feature fusion is not supported. The authors should report mean and standard deviation over multiple random seeds and, ideally, per-dataset breakdowns for the ablation.","section":"III.D, Table II"},{"comment":"Section III.C / Fig. 3: the 'selective feature expression' mechanism is not formally defined. The text states only that each feature is fed into 'corresponding LSTM and softmax layers' and that the SFE-Net 'selectively enhances' features; no equations describe how the selection weights are computed, trained, or applied, nor how the LSTM outputs are combined with the landmark features in the classifier. Without this specification, the paper cannot be replicated and the claim that the module implements dynamic feature selection rather than a standard late-fusion classifier is unverifiable.","section":"III.C, Fig. 3"},{"comment":"Section IV: the conclusion acknowledges 'performance variance on complex datasets, relatively high computational costs, and preprocessing dependence' as limitations, but the experiments contain no measurements of variance, compute, or preprocessing sensitivity. These asserted limitations directly qualify the cross-dataset generalization claim and should be quantified or removed; as written, the text concedes that the robustness conclusion is untested.","section":"IV (Conclusion)"}],"minor_comments":[{"comment":"Figure 3 uses two '(c)' subfigure labels ('(c) Feature Extraction' and '(c) SFE-Net'), and the subfigure labels in the caption do not match the surrounding text; this makes the architecture figure hard to follow.","section":"Fig. 3"},{"comment":"Section III.B says the paper reports video-level AUC, AP, and EER, but Table I contains only frame-level AUC; either add the promised metrics or delete the sentence.","section":"III.B"},{"comment":"Table I lists SFE-Net with backbone 'Xception', but Section III.C does not mention Xception anywhere in the architecture description; the role of Xception (feature extractor, classifier, or both) needs to be stated.","section":"III.C, Table I"},{"comment":"The biological analogy is described at length, but no connection is made to any mathematical or algorithmic design choice; the analogy should be either operationalized or substantially reduced.","section":"I, Fig. 1"},{"comment":"Several sentences are grammatically incomplete or mispunctuated (e.g., 'We show in Fig. 1. how to understand feature selective expression based on gene selective expression.' and 'In the SFE-Net, each of these features is fed into corresponding LSTM and softmax layers to further extract discontinuities between adjacent frames.'); a careful language edit is needed.","section":"Throughout"},{"comment":"The text refers to 'DeepFakeDetection [25]' while Table I uses the abbreviation 'DFDCP'; please make the dataset naming consistent.","section":"III.A, Table I"}],"recommendation":"major_revision","confidential_remarks":"The paper is very thin for a full journal article: the central empirical claim is plausible but currently unverified because the SFE-Net evaluation protocol is not described and no uncertainty quantification is provided. I recommend major revision rather than rejection because the missing information (full protocol, repeated runs, significance testing, architectural equations) is, in principle, obtainable from the authors. If the authors cannot supply the exact evaluation protocol, the cross-dataset superiority claim should be withdrawn."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a plausible incremental combination of known features for deepfake detection, but the central claim—that SFE-Net beats the best cited baseline by about 0.03 average AUC in cross-dataset testing—rests on an unspecified evaluation protocol. I can't tell whether the gain is real or a measurement artifact.\n\nWhat's actually new: the architecture takes five handcrafted forensic features (light consistency, high frequency, compression artifacts, morphological ops, texture), runs each through its own LSTM stream, then uses a learned selection to combine them. That's a reasonable arrangement of existing components. The gene-expression analogy is just framing; it doesn't buy you a formal mechanism. The paper does report an ablation and is honest that the gain is modest and that DFDC remains hard.\n\nThe soft spots are serious. Section III.E says only that the baselines use DeepfakeBench weight files; it never says how SFE-Net was trained or evaluated. No optimizer, learning rate, batch size, frame count, face-crop details, or aggregation rule. Without that, Table I's 0.795 vs 0.767 is not a controlled comparison; face-crop padding and frame selection alone can move AUC by several points. The ablation in Table II has no error bars or repeated runs, and the selection mechanism is described only through a dense figure—no equations for the gating or the LSTM states. There's also no code or weights released. The 0.028 margin over SPSL is exactly the size of protocol sensitivity in this literature.\n\nThat said, the paper isn't incoherent. The method is describable, the datasets are standard, and the results are plausible. The flaw is under-reporting, not fabrication. If the authors can specify the protocol and release code, the claim becomes checkable.\n\nWho it's for: anyone tracking incremental deepfake detection methods or using DeepfakeBench as a baseline. I wouldn't cite it now, but it's a reasonable warning example of why protocol reporting matters. I'd send it to review only with the explicit expectation that the missing details be supplied; in current form, it shouldn't be accepted.","headline":"A plausible but under-specified deepfake detector whose headline cross-dataset gain rests on an unstated evaluation protocol.","tokens_in":7836,"tokens_out":3520,"would_cite":false,"duration_ms":31438,"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":"SFE-Net claims a dynamic feature-selection module lifts deepfake detection to 0.795 average frame-level AUC across five unseen datasets.","keywords":["deepfake detection","feature selection","differential gene expression","cross-dataset generalization","frequency-domain features","dynamic feature weighting","FaceForensics++","LSTM"],"falsifier":"Re-running SFE-Net inside the same evaluation harness that produced the baseline numbers, with identical FF++ c23 training data, cropping, frame sampling, and AUC code, would settle the claim: if the average frame-level AUC over CDF-v1, CDF-v2, DFD, DFDC, and DFDCP falls to 0.767 or below, the reported cross-dataset margin disappears; if it stays near 0.795, the module's contribution is confirmed under the benchmark that generated the baseline results.","tokens_in":6796,"feed_emoji":"🧬","tokens_out":6569,"duration_ms":59755,"temperature":0.7,"pith_summary":"The paper asks whether a deepfake detector can do better by dynamically re-weighting its own feature channels depending on what kind of forgery it sees, rather than relying on static features. It proposes SFE-Net, a Selective Feature Expression Network that borrows the idea of differential gene expression: just as a cell activates different genes in different environments, the network activates different forensic cues for different inputs. On the standard protocol of training on FaceForensics++ (FF++ c23) and testing on five unseen deepfake datasets, SFE-Net reports 0.795 average frame-level AUC, above the best cited baseline SPSL at 0.767. The same module is shown to lift the fusion of five hand-crafted features from 0.771 to 0.795 in an ablation. A sympathetic reader would take the contribution as evidence that explicit, input-dependent feature selection is a viable path to cross-dataset generalization.","feed_headline":"Feature-switching detector hits 0.795 average deepfake AUC","feed_subtitle":"Weighting five forensic cues per input beats static baselines on five unseen deepfake datasets.","key_machinery":"The load-bearing object is the SFE-Net selection module itself, a small network placed after the five feature streams and before the classifier, which maps each input to a distribution of feature priorities, implemented with small convolutional branches, pooling, and softmax weighting. It plays the role of a gene-regulatory switch: the same set of forensic cues is available to the model at all times, but the module decides which cues are 'expressed' for the current frame. Its work is to convert five static feature maps into a context-dependent combination, and the ablation attributes the entire improvement from 0.771 to 0.795 to this module rather than to any one feature.","core_discovery":"On the paper's own terms, the central claim is that inserting a selective feature expression module before a classifier lets a deepfake detector adapt its feature priorities to the synthesis method, and that this dynamic weighting transfers across datasets. The paper defines five forensic features, namely lighting consistency, high-frequency phase, compression reconstruction, morphological edge continuity, and texture, each extracted from preprocessed frames and fed through its own LSTM and softmax stream. SFE-Net then mixes these streams with a selection module, so the relevance of each cue is decided per input. Tables report that this design reaches 0.795 average frame-level AUC on Celeb-DF-v1, Celeb-DF-v2, DFD, DFDC, and DFDCP after training on FF++ c23, compared with 0.767 for the strongest cited baseline, and that removing SFE-Net drops the fusion of all five features to 0.771.","pith_inferences":["If the reported protocol is exactly the one that produced the baseline numbers, namely the same FF++ c23 split, face preprocessing, frame sampling, and evaluation code, then a margin near 0.028 AUC is meaningful evidence that dynamic feature selection helps cross-dataset transfer; the paper never states that its own training and evaluation matched those baselines, so a reader should ask for that p","The biological analogy suggests a testable extension: inspect the learned feature weights on inputs from different synthesis families and check whether SFE-Net upweights phase and high-frequency cues on GAN-based forgeries and edge or morphological cues on warping-based ones.","The same gating idea could be ported to a single learned embedding instead of five hand-crafted features, yielding a test of whether the mechanism works by re-weighting alone rather than by the specific forensic features chosen here.","A stress test on heavily compressed or low-resolution videos would clarify whether the module's dynamic weighting remains useful when high-frequency and phase cues are largely destroyed."],"forward_implications":["Training on FF++ c23, SFE-Net reports frame-level AUC of 0.795 averaged over five unseen datasets, a margin of 0.028 over the best cited baseline SPSL.","Adding the selection module to the five-feature fusion raises average AUC from 0.771 to 0.795, so the paper attributes the gain to the module rather than to feature count.","The approach keeps an Xception backbone and standard FF++ training, meaning the reported improvement comes from changing the head of the pipeline, not from a new pretraining scheme.","On individual datasets the largest gains are on CDF-v1 at 0.866 and CDF-v2 at 0.798, while DFDC remains the weakest point at 0.709, which the authors state as a limitation.","Because the module is feature-agnostic, the same gating idea could be applied to other feature extractors without retraining the backbone from scratch."],"supporting_citations":[{"why":"Supplies the FaceForensics++ c23 training corpus and the standard training protocol for all detectors in the comparison.","marker":"[10]"},{"why":"Provides the pre-trained weight files whose evaluations produce the baseline AUC numbers in Table I, anchoring the comparison to that benchmark.","marker":"[26]"},{"why":"Provides the Celeb-DF-v1 and Celeb-DF-v2 evaluation sets used for two of the five cross-dataset test columns.","marker":"[22]"},{"why":"Contributes the DFDCP evaluation set used as a third cross-dataset test column.","marker":"[23]"},{"why":"Contributes the DFDC evaluation set, the column where SFE-Net scores lowest.","marker":"[24]"},{"why":"Supplies the DeepFakeDetection evaluation set, the fifth cross-dataset test column.","marker":"[25]"},{"why":"Defines SPSL, the frequency-domain baseline with the strongest reported average AUC that SFE-Net is claimed to surpass.","marker":"[20]"}],"fun_headline_variants":["Adaptive deepfake detector switches cues per input for 0.795 AUC","Bio-inspired feature selection boosts deepfake AUC to 0.795","SFE-Net tunes five cues per frame for 0.795 deepfake AUC","Dynamic cue prioritization yields 0.795 AUC on deepfake sets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that SFE-Net was trained and evaluated under exactly the same protocol as the baselines, meaning the same FF++ c23 split, face preprocessing, frame sampling, and evaluation code, since the paper borrows baseline weight files from a benchmark but does not describe its own training details.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive deepfake detector switches cues per input for 0.795 AUC","Bio-inspired feature selection boosts deepfake AUC to 0.795","SFE-Net tunes five cues per frame for 0.795 deepfake AUC","Dynamic cue prioritization yields 0.795 AUC on deepfake sets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000829,"raw_usage":{"total_tokens":3639,"prompt_tokens":979,"completion_tokens":2660,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":595,"completion_tokens_details":{"reasoning_tokens":2578}},"tokens_in":595,"tokens_out":2660,"duration_ms":17550,"temperature":1.0,"reasoning_tokens":2578,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:11:14.550970+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-running SFE-Net inside the same evaluation harness that produced the baseline numbers, with identical FF++ c23 training data, cropping, frame sampling, and AUC code, would settle the claim: if the average frame-level AUC over CDF-v1, CDF-v2, DFD, DFDC, and DFDCP falls to 0.767 or below, the reported cross-dataset margin disappears; if it stays near 0.795, the module's contribution is confirmed under the benchmark that generated the baseline results.","supporting_citations":[{"cited_title":"Rossler, D","cited_arxiv_id":null,"evidence_quote":"Supplies the FaceForensics++ c23 training corpus and the standard training protocol for all detectors in the comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the pre-trained weight files whose evaluations produce the baseline AUC numbers in Table I, anchoring the comparison to that benchmark."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Celeb-DF-v1 and Celeb-DF-v2 evaluation sets used for two of the five cross-dataset test columns."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the DeepFakeDetection evaluation set, the fifth cross-dataset test column."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines SPSL, the frequency-domain baseline with the strongest reported average AUC that SFE-Net is claimed to surpass."}],"review_version":1}