{"id":"3ec507c9-3e11-41b2-8731-4b7b2d5cca61","arxiv_id":"2507.15101","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Partial deepfake speech is detected by scoring unnatural frame-to-frame changes in self-supervised audio embeddings, reaching 0.59% EER on PartialSpoof and 0.03% on HAD with utterance-level labels only.","lead":"This paper introduces a deepfake speech detector that looks at how audio features change from one moment to the next, instead of hunting for the seam where fake audio is spliced in. It reports record-low error rates on two partial-deepfake benchmarks while training only on utterance-level labels, not expensive frame-by-frame labels.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The temporal-difference cue is validated only on fully fake LA utterances, and no duration-stratified results show it survives pooling for short PartialSpoof segments; this leaves the SOTA generalization claim conditional.","rationale":"I read the paper in good faith. The proposed TDAM is coherent, the ablations show that each design choice contributes, and the code is publicly available, which are real supporting elements. The central concern is not that the reported numbers are fabricated but that the mechanism asserted to produce them is validated only indirectly. Section III motivates the whole approach with boxplots on fully fake ASVspoof2019-LA utterances and provides no quantitative measure of separation, no segment-level analysis on partial spoofs, and no duration-stratified evaluation. The reader's weakest assumption correctly identified this gap: adaptive average pooling to a fixed temporal length is only safe if the directional cue survives for short injected segments, and the paper does not demonstrate that. I also note the ambiguity in the description of T', which makes it difficult to know the effective temporal resolution after pooling. Because the claim explicitly includes generalization to unseen attacks and operation without frame-level supervision, a duration-stratified test is necessary to support the headline result. This does not overturn the paper's results, but it confirms that the verdict should remain conditional on that missing evidence.","tokens_in":8420,"tokens_out":6630,"duration_ms":75852,"concrete_test":"Using the public code and PartialSpoof ground-truth segment labels (for evaluation only, not training), stratify the evaluation set by manipulated-segment duration (<0.5 s, 0.5-1 s, 1-2 s, >2 s) and report TDAM-AvgPool EER per stratum with bootstrap 95% confidence intervals. Separately, recompute the Eqs. (1)-(2) mean and standard deviation of cosine similarity on fake-only versus bonafide-only frames within each stratum. If the short-stratum EER is materially worse than the overall 0.59% or the mu/sigma separation disappears for short fake segments, the central claim that temporal-difference cues generalize to short partial deepfakes is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is that first-order directional temporal differences in wav2vec2-XLSR embeddings are a generalizable cue for synthetic segments, and that the pooling scheme in Eq. (3) preserves this cue when injected fake segments are short. Section III supports this premise only with unquantified boxplots on ASVspoof2019-LA, which contains fully fake utterances, not partial spoofs: no numbers, confidence intervals, or effect sizes are reported, and no analysis separates fake frames from surrounding bonafide frames in PartialSpoof. The implemented TDAM difference map in Eq. (4) also uses raw unnormalized differences after a temporal convolution, not the normalized directions and cosine similarities of Eqs. (1)-(2), so the motivating statistic is not exactly what the model consumes. The pooling description is ambiguous: V-B says \"T' is set to 4 seconds\" even though T' is an integer number of segments in Eq. (3). Depending on whether each pooled output is about 20 ms or about 1 s, a short fake segment is either preserved or averaged away. Critically, no duration-stratified results are reported for PartialSpoof, whose evaluation set contains unseen attacks and short embedded segments. If the directional cue separates only whole utterances or long fake regions, the headline EERs would not transfer to the short-segment regime that defines partial deepfake detection. The cross-dataset table also shows TDAM is worse than a strong baseline on LA when trained on LA (1.22 vs 0.83 EER), so the evidence for a universal marker is mixed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a weakly supervised method for partial deepfake speech detection. It uses wav2vec2-XLSR frame embeddings, computes frame-to-frame difference features, and introduces a Temporal Difference Attention Module (TDAM) with a dual-level hierarchical representation to capture temporal irregularities without frame-level labels. Variable-length inputs are handled by adaptive average pooling. The authors report state-of-the-art results on PartialSpoof (EER 0.59%) and Half-Truth (HAD, EER 0.03%) and provide ablations showing that TDAM, the dual-level structure, directional information, and average pooling all contribute.","tokens_in":8704,"tokens_out":4116,"duration_ms":45318,"significance":"If the results hold, the paper offers a useful advance: it removes the need for frame-level supervision while reporting strong accuracy on partial deepfake benchmarks, and it introduces a temporal-difference perspective that is well motivated by an analysis of wav2vec2-XLSR embeddings. The public code and the internally consistent ablation study are strengths. However, the headline claim of state-of-the-art performance is currently supported only by single-run EERs with very small margins over strong baselines, and the connection between the motivating normalized-direction statistics and the actual raw-difference module is not quantitatively established. The significance is therefore conditional on additional statistical and generalization evidence, especially for short fake segments.","major_comments":[{"comment":"All reported EERs are single-run values with no error bars, seeds, or statistical significance tests. This is load-bearing because the claimed advantage over the strongest baseline is very small on PartialSpoof (0.59% vs. 0.64% for Zhang et al. in Table I) and 0.03% vs. 0.06% on HAD. The abstract and Section VI-A say the method 'significantly outperforms' existing methods, but the paper provides no evidence of significance. Please report mean and standard deviation over multiple runs (or at least fixed seeds) and, where appropriate, a paired test.","section":"Section VI-A, Tables I and II"},{"comment":"The motivating analysis in Section III uses normalized direction vectors and cosine similarities (Eqs. (1)-(2)), but the difference map actually implemented in Eq. (4) is a raw, unnormalized difference Econv[t+1]-E[t] after a temporal convolution. No experiment demonstrates that the normalized-direction statistic separates partial fake frames from surrounding bonafide frames, nor that the model's performance is attributable to that statistic rather than to magnitude-based differences. Please quantify the motivating statistic on the PartialSpoof training/evaluation data (e.g., effect sizes or distributions) and clarify how Eq. (4) realizes the directional cue.","section":"Section III and Eq. (4)"},{"comment":"The temporal-difference cue is validated only on ASVspoof2019-LA, which consists of fully fake utterances, yet the paper's central claim concerns partial deepfakes with short embedded fake segments. No duration-stratified results are reported for the PartialSpoof evaluation set. Because the evaluation set contains unseen attacks and short injected segments, the reader cannot tell whether the cue survives the pooling operation when the fake region is brief. Please report EER broken down by fake-segment duration (or by proportion of fake frames), and show corresponding frame-level localization quality, to support the claimed generalization.","section":"Section III and Section VI"},{"comment":"The notation for the pooling target is inconsistent: Eq. (3) defines T' as an integer number of pooled segments, while Section V-B says 'T' is set to 4 seconds.' Since wav2vec2-XLSR embeddings are extracted every 20 ms, T'=4 seconds is not a well-defined number of segments. The effective pooled segment duration determines whether a short fake segment is preserved or averaged away, so this is not merely a wording issue. Please define T' precisely, specify the resulting segment duration, and report sensitivity of the PS evaluation EER to T'.","section":"Eq. (3) and Section V-B"},{"comment":"The text states that 'our model outperforms state-of-the-art models' in the cross-dataset study, but in the 'Training on LA' condition the proposed method has a higher EER on LA eval (1.22%) than Zhang et al. (0.83%). The claim is therefore not true as written. Please revise the claim and discuss what this weaker cross-dataset result implies for the 'strong generalizability' conclusion, especially given that the model is worse on fully fake utterances than a strong baseline when trained on LA.","section":"Table IV and Section VI-C"}],"minor_comments":[{"comment":"The symbol l in Eq. (5) is used for a 2D convolution but visually resembles the number 1 and the index variable; please use a different notation, such as W or K, to avoid confusion.","section":"Section IV-B, Eq. (5)"},{"comment":"The description of X1 and X2 after the 2D convolutions is incomplete: the kernel sizes, padding, strides, and how the output shape R^{C x T' x D} is obtained are not specified. Adding these details would improve reproducibility, even though the code is available.","section":"Section IV-B"},{"comment":"The boxplots in Figure 2 are described only qualitatively. Please report the number of utterances, the means, standard deviations, and an effect-size measure so that the claimed separation between bonafide and deepfake cosine-similarity distributions is supported by numbers rather than visual inspection.","section":"Figure 2"},{"comment":"The sentence 'Based on duration statistics of the training datasets, we set T' to 4 seconds, matching the average utterance length' is imprecise because Eq. (3) treats T' as a count of segments. Please rephrase to state the actual number of segments used.","section":"Section V-B"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is in scope for the journal and the core idea is interesting, but the load-bearing evidence needs strengthening before the SOTA claim can be accepted. The most important additions are multiple-run statistics, duration-stratified results on PartialSpoof, and a clearer alignment between the motivating statistic and the implemented module. I do not see grounds for rejection because the ablations are consistent and the issues are addressable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: the paper's real contribution is a new weakly supervised angle on partial deepfake detection — frame-level temporal difference attention, no boundary or frame labels needed — and the ablations back up the module design. That part is solid and worth building on. But the headline \"state-of-the-art, significantly outperforms\" claim is under-supported by the evidence as reported.\n\nWhat's genuinely new: TDAM's dual-level difference representation, coupled with adaptive average pooling for variable-length inputs, is a sensible and clean idea. The motivation from video deepfake temporal inconsistency is honestly credited. The ablation table shows each component matters: removing the whole module, removing the dual-level structure, or removing directional information all degrade EER. That is real evidence of internal validity.\n\nWhere it gets soft. First, all EERs are single-run values with no error bars, seeds, or significance testing. The PartialSpoof gain over the strongest published baseline (Zhang et al., 0.64 EER) is 0.59 — a 0.05 point difference, which could easily be noise. Second, the motivating analysis (Sec. III) uses normalized direction vectors and cosine similarity, but the implemented difference map in Eq. (4) uses raw differences after a 1D convolution. That gap is never discussed. Third, the pooling description is ambiguous: Sec. V-B says \"T' is set to 4 seconds\" but Eq. (3) treats T' as an integer number of segments. If each pooled output spans about a second, a short fake segment could be averaged away; if it spans 20 ms, the cue survives. No duration-stratified results are reported for PartialSpoof, whose whole point is short embedded fake segments. That is a load-bearing gap. Fourth, the cross-dataset table shows TDAM trained on LA gets 1.22 EER on LA eval, worse than Zhang et al.'s 0.83. If the temporal-difference cue were a universal marker, you'd expect it to at least match on fully fake utterances.\n\nThe paper is coherent, the writing is clear, and I don't see circular benchmarking or load-bearing self-citation. The central idea is credible; the evidence just isn't complete enough to establish the strong generalization claim.\n\nWho this is for: the audio anti-spoofing subfield, especially people working on partial deepfakes and weakly supervised localization. I'd bring it to reading group and would cite it once the numbers are confirmed. A serious editor should send it to peer review — the idea and the ablations deserve referee time — but the authors should be pushed to add uncertainty quantification, clarify the pooling semantics, and report duration-stratified results before the SOTA claim is taken at face value.\n\nRecommendation: engage, but treat the headline numbers as provisional.","headline":"A worthwhile weakly-supervised partial deepfake detection idea with solid ablations, but the headline generalization claim outruns the evidence because the short-segment regime is never directly tested.","tokens_in":9234,"tokens_out":1531,"would_cite":true,"duration_ms":18979,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Partially faked speech can be detected from erratic frame-direction changes in acoustic embeddings, using only utterance-level labels; the reported equal error rates are 0.59% on PartialSpoof and 0.03% on Half-Truth.","keywords":["partial deepfake speech detection","temporal difference learning","wav2vec2-XLSR","weakly supervised detection","adaptive average pooling","anti-spoofing","temporal attention","cross-lingual generalization"],"falsifier":"Sort PartialSpoof evaluation utterances into bins by the total duration of injected deepfake segments and recompute equal error rate per bin: if the shortest-duration bin shows a large EER increase, the claim that pooling preserves short-segment directional cues fails.","tokens_in":8209,"feed_emoji":"🎙️","tokens_out":6764,"duration_ms":65379,"temperature":0.7,"pith_summary":"Partially faked speech—utterances where only some segments are synthetic—can be detected without expensive frame-level labels. The paper shows that in wav2vec2-XLSR audio embeddings, bonafide speech moves with smoother, more consistent frame-to-frame direction, while deepfake segments fluctuate erratically. It introduces a Temporal Difference Attention Module that learns to weight frames by these irregularities at fine and coarse time scales, then averages frame predictions into an utterance verdict. On the PartialSpoof and Half-Truth datasets the model reaches equal error rates of 0.59% and 0.03%, respectively, outperforming systems that rely on transition boundaries or segment labels.","feed_headline":"Temporal differences catch partial deepfakes without frame labels","feed_subtitle":"A weakly supervised model hits 0.59% EER on PartialSpoof and 0.03% on HAD, surpassing boundary-based baselines.","key_machinery":"The load-bearing component is the Temporal Difference Attention Module (TDAM), which operates on a refined embedding $E$ of $T'$ frames. It first computes the directional difference map $M[t] = E_{\\text{conv}}[t+1] - E[t]$ for $t = 1,\\dots,T'-1$ with a zero vector appended to keep the shape. This map feeds two convolutional levels: $X_1$ catches fine-grained local inconsistencies, and $X_2$ is formed by downsampling, convolving, and upsampling to capture longer-term motion. The sum $X_1 + X_2$ passes through a sigmoid to yield frame confidence scores, and the output is $Y = \\sigma[l*(X_1+X_2)] \\odot E$, where $l$ is a $1\\times1$ convolution. A softmax over $Y$ gives frame predictions that are averaged into the utterance verdict. Adaptive average pooling first partitions the variable-length frames into $T'=4$ second segments so the pipeline accepts any input duration.","core_discovery":"The central claim is that frame-level temporal difference direction is a sufficient cue for partial deepfake detection when computed on wav2vec2-XLSR embeddings. In the paper's motivating analysis, bonafide utterances yield higher mean and lower standard deviation of the cosine similarity between adjacent normalized direction vectors, indicating natural prosodic flow, whereas deepfakes show oscillatory directional changes. On this basis, TDAM-AvgPool captures both fine-grained and long-range temporal irregularities and, trained only on utterance-level labels, reports equal error rates of 0.19% on the PartialSpoof development set, 0.59% on its evaluation set, and 0.03% on the Half-Truth test set, plus 99.99% AUC on HAD. The ablations attribute most of the gain to the TDAM module itself and to preserving directional rather than absolute differences.","pith_inferences":["Because the temporal-difference cue is defined on any frame embedding sequence, the same TDAM pipeline could be tested on other self-supervised encoders and on streaming speech where frame differences are computed causally.","The boxplot separation in Figs. 1 and 2 is the paper's motivating evidence but is not quantified; a natural extension is to report effect sizes or train a simple classifier on just the mean and standard deviation statistics to isolate how much of the EER comes from the statistics versus the learned attention.","If the directional-smoothness cue degrades for very short injected segments, duration-stratified EER would expose a practical lower bound on manipulation length the method can catch, a limit the current evaluation does not address."],"forward_implications":["Frame-level annotations are no longer needed to train a partial deepfake detector, removing a major scalability bottleneck.","Smoothing of transition artifacts no longer breaks detection, because the model listens for irregularities inside manipulated regions rather than at boundaries.","The same detector works on fully fake utterances, since it learns artifact patterns rather than boundary patterns.","English and Chinese partial deepfakes are handled by one model, pointing toward cross-lingual deployment without dataset-specific tuning.","Ablations show that removing TDAM raises the PartialSpoof evaluation EER from 0.59% to 0.87%, so the temporal-difference cue carries the performance, not the SSL front-end alone."],"supporting_citations":[{"why":"PartialSpoof database and countermeasures; sets the task and the evaluation set with 11 unseen attacks, including smoothing-based generation methods.","marker":"[6]"},{"why":"ASVspoof2019-LA training set; used for the boxplot analyses of frame-difference statistics that motivate the temporal-difference cue.","marker":"[25]"},{"why":"wav2vec2-XLSR; the pre-trained SSL front-end that provides the 1024-dimensional frame embeddings on which all differences are computed.","marker":"[26]"},{"why":"Half-Truth (HAD) dataset; provides the Chinese, accent-varied test set that supports the cross-lingual generalization claim.","marker":"[27]"},{"why":"Cai and Li; a boundary-integration baseline from which the proposed method's boundary-free design is contrasted and which it outperforms on both PS and HAD.","marker":"[22]"},{"why":"Liu et al.; a study of how countermeasures localize partial spoofing, used as a comparison system on the PS evaluation set.","marker":"[23]"},{"why":"Gu et al.; spatiotemporal inconsistency learning for video deepfakes, the cross-domain motivation for analyzing temporal irregularities.","marker":"[24]"}],"fun_headline_variants":["Temporal differences spot partial deepfakes without frame labels","Frame-level temporal difference learning detects partial deepfakes","No boundary labels needed: temporal differences uncover partial deepfakes","Temporal irregularity cues reveal partial deepfakes, no frame labels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that first-order directional irregularity in wav2vec2-XLSR frame embeddings is a generalizable marker of synthetic speech, and that adaptive average pooling to 4-second segments preserves this marker even when the injected fake segments are short.","fun_headline_variants_meta":{"raw":{"variants":["Temporal differences spot partial deepfakes without frame labels","Frame-level temporal difference learning detects partial deepfakes","No boundary labels needed: temporal differences uncover partial deepfakes","Temporal irregularity cues reveal partial deepfakes, no frame labels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000525,"raw_usage":{"total_tokens":2528,"prompt_tokens":928,"completion_tokens":1600,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":1530}},"tokens_in":544,"tokens_out":1600,"duration_ms":11786,"temperature":1.0,"reasoning_tokens":1530,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:40:50.437004+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Sort PartialSpoof evaluation utterances into bins by the total duration of injected deepfake segments and recompute equal error rate per bin: if the shortest-duration bin shows a large EER increase, the claim that pooling preserves short-segment directional cues fails.","supporting_citations":[{"cited_title":"The partialspoof database and countermeasures for the detection of short fake speech segments embedded in an utterance,","cited_arxiv_id":null,"evidence_quote":"PartialSpoof database and countermeasures; sets the task and the evaluation set with 11 unseen attacks, including smoothing-based generation methods."},{"cited_title":"Asvspoof 2019: Future horizons in spoofed and fake audio detection,","cited_arxiv_id":null,"evidence_quote":"ASVspoof2019-LA training set; used for the boxplot analyses of frame-difference statistics that motivate the temporal-difference cue."},{"cited_title":"Xls-r: Self-supervised cross-lingual speech representation learning at scale,","cited_arxiv_id":null,"evidence_quote":"wav2vec2-XLSR; the pre-trained SSL front-end that provides the 1024-dimensional frame embeddings on which all differences are computed."},{"cited_title":"Half- truth: A partially fake audio detection dataset,","cited_arxiv_id":null,"evidence_quote":"Half-Truth (HAD) dataset; provides the Chinese, accent-varied test set that supports the cross-lingual generalization claim."},{"cited_title":"Integrating frame-level boundary detection and deepfake detection for locating manipulated regions in partially spoofed audio forgery attacks,","cited_arxiv_id":null,"evidence_quote":"Cai and Li; a boundary-integration baseline from which the proposed method's boundary-free design is contrasted and which it outperforms on both PS and HAD."},{"cited_title":"How do neural spoofing countermeasures detect partially spoofed audio?","cited_arxiv_id":null,"evidence_quote":"Liu et al.; a study of how countermeasures localize partial spoofing, used as a comparison system on the PS evaluation set."},{"cited_title":"Spatiotemporal inconsistency learning for deepfake video detection,","cited_arxiv_id":null,"evidence_quote":"Gu et al.; spatiotemporal inconsistency learning for video deepfakes, the cross-domain motivation for analyzing temporal irregularities."}],"review_version":1}