{"id":"d227429c-159a-4a15-bee8-0a2c8c92f2f9","arxiv_id":"2607.16873","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"InfoDense replays only density-ranked, forgery-decisive face fragments rather than full images, cutting memory use and improving incremental deepfake detection.","lead":"InfoDense is a replay method for incremental deepfake detection that stores only small, artifact-dense facial fragments instead of full images, then stitches them into new training samples. It reports major gains in accuracy and cross-dataset generalization at roughly 80% lower storage, but the comparisons as presented are not fully controlled.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported SOTA gains are not controlled for backbone or replay memory budget, and key table rows contain copy errors; the central claim is unsupported as presented.","rationale":"The reader's weakest_assumption identifies the exact load-bearing concern: the submitted comparison is not controlled for backbone or replay-memory budget. My reading of the full text confirms this and adds concrete evidence: the implementation section reveals a CLIP ViT-L/14 backbone while baseline rows are reproduced from papers that use other backbones; no m or n values are given for Table 1; and the ablation tables contain apparent copy errors (Table 2 Base(DFIL) cross-dataset equals L-Bound; Table 3 Random equals Table 2 ClockMix). These are not merely stylistic issues—they prevent a reader from verifying whether the reported gains are due to the proposed regional decisive replay strategy or to confounding factors. The central claim of the paper is that InfoDense simultaneously improves memory efficiency, catastrophic forgetting, and cross-domain generalization; the evidence for that claim is the comparison table. Since the table is uncontrolled and internally inconsistent, the paper is not acceptable as-is. However, the method itself is plausible, and the internal ablations (though flawed) suggest the regional selection may help. Therefore the appropriate disposition is a conditional acceptance requiring matched reimplementation and correction of the table errors, not outright rejection. My recommendation is UNCHANGED relative to the reader's CONDITIONAL verdict.","tokens_in":24509,"tokens_out":5894,"duration_ms":59206,"concrete_test":"Re-run Protocol 1 with all baselines (DFIL, SUR-LID, GPL, and a no-replay lower bound) implemented on the same frozen CLIP ViT-L/14 backbone, fixing the replay memory budget in bytes (choose m and n so total storage is equal, or fix total bytes and let m vary), and report the corrected Table 2 Base(DFIL) cross-dataset row. If the InfoDense margins over baselines shrink below ~1 AUC point or disappear, the central claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim that InfoDense outperforms prior IFFD methods rests on Table 1, but that table compares InfoDense (pre-trained CLIP ViT-L/14 backbone) against published numbers of baselines that were obtained with different architectures, and no replay memory budget or compression ratio n is specified for any row. The text states the same-memory-budget argument, but the table cannot be audited: without knowing how many samples each baseline stored and at what resolution, the reported gains (e.g., +4.13 AUC over SUR-LID on incremental, +7.42 over GPL on cross-dataset) could stem from the stronger backbone or from a much larger effective replay count, not from regional decisive replay. Moreover, the ablation tables contain apparent copy errors: Table 2's Base(DFIL) cross-dataset row (43.36, 88.31, 51.90, 63.87, 61.86) is identical to the L-Bound row in Table 1, not to DFIL's own cross-dataset row (76.53, 84.87, 70.01, 87.03, 79.61); Table 3's Random row shares its entire cross-dataset row with Table 2's ClockMix. These inconsistencies make even the internal ablations unreliable. Because the headline comparison is the load-bearing evidence for the central claim, and it is neither controlled nor internally consistent, the claim cannot be accepted as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes InfoDense, a replay-based incremental face forgery detection (IFFD) method that stores only compact, forgery-decisive facial regions instead of full images. The pipeline consists of InfoDense Cut (CLIP-based localization of decisive patches and extraction of the fan-shaped segment containing the most such patches), InfoDense Select (ranking candidates by a weighted combination of latent-space representativeness and decisive-patch density), and InfoFuse (reconstructing training inputs by fusing stored segments with current-task samples, including inter-task and intra-task mixing). The authors report state-of-the-art results on a Protocol 1 benchmark: incremental average AUC 90.42 vs. 86.29 for SUR-LID, and cross-dataset average AUC 90.03 vs. 82.61 for GPL, while claiming substantial storage reduction and privacy benefits. The paper also includes ablations on compression strategies, selection strategies, fusion strategies, and storage-performance trade-offs.","tokens_in":24885,"tokens_out":6139,"duration_ms":59301,"significance":"The core idea — that regional decisive replay can reduce memory, increase replay diversity, and simultaneously improve cross-domain generalization — is well-motivated and potentially impactful for continual deepfake detection. The privacy analysis with ArcFace/CosFace retrieval is a useful and notable addition, and the proposed selection criterion in Eq. (4) is concrete and falsifiable. If the empirical claims are confirmed under controlled conditions, InfoDense would be a meaningful step toward memory-efficient and privacy-aware IFFD. However, the central empirical claim is currently not established: the main comparison is not controlled for backbone or replay budget, and the internal ablation tables contain apparent copy errors. These issues are load-bearing and must be addressed before the contribution can be assessed.","major_comments":[{"comment":"The headline comparison is not controlled. Implementation Details (§3.1) state that InfoDense uses a pre-trained CLIP ViT-L/14 backbone, but the table gives no information about the backbone, replay memory budget, or buffer size m for any baseline row. The text claims 'under the same memory budget,' but the budget is never specified. Because the baseline rows are taken from published papers (e.g., DFIL MM'23, SUR-LID CVPR'25) that use different architectures, the reported gains (+4.13 incremental, +7.42 cross-dataset vs. the best baselines) could reflect the stronger backbone or a larger replay count rather than the regional replay strategy. Please re-run all baselines on the same CLIP ViT-L/14 backbone with identical replay storage budget, training epochs, and optimizer, and report the budget explicitly.","section":"§3.2, Table 1"},{"comment":"The ablation tables contain apparent copy errors that make the internal evidence unreliable. In Table 2, the Base(DFIL) cross-dataset row (43.36, 88.31, 51.90, 63.87, 61.86) is identical to the L-Bound row of Table 1, not to DFIL's own row in Table 1 (76.53, 84.87, 70.01, 87.03, 79.61). In Table 3, the Random cross-dataset row is identical to the ClockMix row of Table 2 (83.40, 84.56, 73.78, 81.20, 80.58). These inconsistencies prevent the reader from verifying the claimed advantages of density-aware selection over random selection and of InfoDense over the DFIL baseline. Please correct the tables and re-verify all numbers.","section":"§3.3, Tables 2 and 3"},{"comment":"The decisive-patch scorer (Eq. (1)) and the representativeness term (Eq. (4)) both use the same CLIP-based model f(·, θ_{t-1}) that is being incrementally trained. Appendix 7.2 confirms that the score is only used after the model has learned forgery-awareness from prior tasks, so the replay selection is not independent of the model's current inductive bias. This creates a self-reinforcing loop: the model selects fragments it already considers decisive, which may amplify dataset-specific artifacts rather than intrinsic forgery cues. The cross-domain generalization claim would be substantially strengthened by comparing against an independent localizer (e.g., zero-shot CLIP or a fixed Grad-CAM from a pre-trained backbone) and showing that the gains persist. Please add such an experiment or otherwise control for this effect.","section":"§2.2, Eq. (1)–(4); Appendix 7.2"},{"comment":"The memory-efficiency claim is expressed only as a pixel-compression ratio n, but actual storage cost depends on the image encoding. Appendix 2 proposes a row-array encoding to make storage reduction 'strictly match' n%, but it is not stated whether the experiments in Fig. 4 and Tables 1–4 use this encoding or simply store masked/cropped images. Furthermore, the replay buffer size m is never specified in the main text. To make the efficiency claim auditable, please report actual storage in bytes (or the number of stored fragments per task) and state m for InfoDense and all baselines.","section":"§3.1, Fig. 4; Appendix 2"}],"minor_comments":[{"comment":"The method is called 'InfoDense Fuse' in the abstract and Section 1, but 'InfoFuse' in Section 2.3. Please use one name consistently.","section":"Throughout"},{"comment":"The definition PD = M0 − MN is ambiguous: M0 and MN are described as average AUC in the base and final sessions, but it is unclear which datasets are included in these averages. Please define explicitly.","section":"§3.1, PD definition"},{"comment":"Table 2 is captioned 'Ablation study on compression strategies,' but the paragraph 'Impact of Regional Shapes and Patch Selection' refers to Table 2 for CutMix vs. ClockMix and CutMix+PD. Either the table or the paragraph should be retitled to avoid confusion.","section":"§3.3, Table 2 caption"},{"comment":"There are stray '/uni0000...' strings in the manuscript (e.g., after Table 4 and in the appendix figure captions), which appear to be corrupted Unicode/PDF artifacts. These need to be removed in the production version.","section":"Appendix and text"},{"comment":"No statement is provided about code or trained model release. For a paper whose central contribution is an algorithmic pipeline, a code-availability statement would be valuable.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The core idea is interesting and potentially publishable, but the current version has two serious problems: the main comparison is not controlled for backbone or memory budget, and the ablation tables contain what look like copy-paste errors. Both are fixable in a major revision, but the experiments need to be re-run and reported with complete settings. I recommend major revision rather than rejection, provided the authors can supply controlled comparisons and correct tables."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: this is a plausible and well-motivated replay compression method for incremental face forgery detection, but the headline numbers should not be believed as reported. The paper stores only the most forgery-relevant fan-shaped segments of each face, selected with CLIP-based decisive-patch scoring plus latent representativeness, and fuses them into current-task images for replay. The motivation that full faces are redundant for replay is sensible, and the appendix makes a good-faith attempt to show that the pixel reduction translates into real byte savings.\n\nWhat is actually new is the combination: CLIP-based patch scoring for localization, fan-shaped density-ranked segment selection, and cut-mix-style fusion all in one IFFD pipeline. I do not see that exact assembly in the cited prior work. Where the ablations are legible, they support the design choices: density-aware selection beats random and center selection, and the fusion strategy beats isolated self-fusion. The privacy leakage analysis is a useful addition.\n\nThe soft spots are real and load-bearing. Table 1 compares against published baseline numbers, not re-implementations on the same CLIP ViT-L/14 backbone, and it reports no replay memory budget or compression ratio for any row. The claimed gains over SUR-LID and GPL could therefore come from the stronger backbone or from storing more samples, not from regional replay. Worse, the copy errors are verifiable: in Table 2, Base(DFIL)'s cross-dataset row is identical to L-Bound's row from Table 1, not to DFIL's own cross-dataset row; in Table 3, Random's cross-dataset row is identical to ClockMix's row in Table 2. Those errors make the internal ablations unreliable as printed. No error bars and no code round out the auditability problem.\n\nNone of this kills the concept. The appendix directly addresses the circularity worry by noting the CLIP scorer is only used in later tasks after the model has some forgery awareness, which is acceptable as a replay-selection prior. The authors seem to know what the right experiments are; they just have not run the controlled version of the main comparison.\n\nFor peer review: yes, send it to a serious referee. The idea deserves evaluation, and the fixes are incremental rather than conceptual. I would not cite the numbers until they are reproduced under matched conditions, but I might cite the method idea. Worth a reading-group slot for a discussion of how much reproducibility matters in continual learning papers.","headline":"The core idea is sound and interesting, but the reported SOTA comparisons are not auditable and two ablation tables contain copy errors; treat the numbers as unverified and ask for a re-run.","tokens_in":25354,"tokens_out":2658,"would_cite":false,"duration_ms":27483,"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":"Replaying only the artifact-dense regions of a face, not the whole image, reduces memory and improves incremental deepfake detection while also cutting identity leakage from the replay buffer.","keywords":["incremental face forgery detection","replay strategy","catastrophic forgetting","cross-domain generalization","CLIP","decisive patches","memory-efficient replay","privacy-preserving replay"],"falsifier":"Re-run the leading baselines (e.g., DFIL, SUR-LID) on the same CLIP ViT-L/14 backbone with the same replay-buffer byte budget and sample count; if their average AUC reaches or exceeds InfoDense's, the regional-selection strategy is not the source of the gain. Alternatively, swap the CLIP-similarity decisive-patch scorer for random scores while keeping the same cut-and-fuse pipeline; a non-drop would show density-awareness is not doing the work.","tokens_in":24414,"feed_emoji":"🎭","tokens_out":7763,"duration_ms":70123,"temperature":0.7,"pith_summary":"The paper argues that face-forgery replay buffers waste limited memory on redundant whole-face images, and that storing only the artifact-dense regions of each face is enough to preserve the knowledge a detector needs. It proposes InfoDense, a three-stage strategy: it cuts each historical face into fan-shaped segments, scores every patch by its similarity to a forgery-representative CLIP text embedding, and stores the segment that concentrates the most decisive patches while remaining representative in latent space. Stored fragments are later fused onto current-task samples, with deliberate inter-task and intra-task mixing to avoid class and task bias. On two incremental deepfake protocols, InfoDense reports higher average AUC than recent replay baselines, a lower forgetting rate, and near-zero identity-retrieval recall for stored samples. A sympathetic reader will take this as evidence that spatial compression of replay data can simultaneously reduce memory, improve generalization, and mitigate a privacy risk.","feed_headline":"Face fragments beat full-face replay in incremental deepfake detection","feed_subtitle":"Storing only artifact-dense regions keeps more history under the same memory, cutting forgetting and identity leakage.","key_machinery":"The carrying mechanism is the 'decisive patch' — a local image region whose CLIP embedding aligns most closely with forgery semantics. Around it, InfoDense builds three operations: InfoDense Cut uses a two-pointer scan over ClockMix fan-shaped segments to find the segment that covers the maximum number of decisive patches; InfoDense Select ranks segments by a weighted combination of latent-space centroid distance and decisive-patch count; InfoDense Fuse reconstructs unbiased training inputs by masking stored segments onto current-task images with inter-task and intra-task mixing. The underlying assumption is that forgery cues are redundant enough across the face that one compact, center-anch","core_discovery":"InfoDense's central claim is that facial forgery evidence is spatially concentrated and center-symmetric, so replaying an entire face is wasteful. The paper retains, for each historical face, only the fan-shaped segment with the largest number of 'decisive patches' — the top-10 CLIP image tokens most similar to a forgery-representative text embedding. Candidate segments are ranked by a weighted sum of cosine similarity to the task's feature centroid (representativeness) and decisive-patch count (information density), and the top-m are stored. At training time, stored segments are pasted onto current-task images with a binary mask, using four real/fake label combinations across tasks plus int","pith_inferences":["The decisive-patch scoring depends on CLIP's text-image alignment and on the encoder having acquired some forgery awareness from earlier tasks; a natural test is whether using the detector's own attention or a non-CLIP local-artifact score changes the gains, which would separate the density prior from the backbone.","The main tables compare InfoDense's CLIP ViT-L/14 results against baselines reproduced from other architectures, so the incremental and cross-domain gaps may shrink if every method is re-run on the same backbone and the same byte budget; a direct re-implementation would settle how much of the gain is the regional-decisive strategy itself.","If the facial-redundancy hypothesis is right, similar regional-decisive replay should benefit other continual perception tasks with localized cues (e.g., forgery of documents, medical image artifacts), where full-image replay is also memory-inefficient.","The privacy argument is measured only by face-recognition retrieval; membership inference or attribute inference on fused samples could still leak identity, so the de-identification claim is narrower than 'privacy-safe'."],"forward_implications":["Under a fixed memory budget, storing 1/n-size segments enables roughly n times more historical samples to be replayed, which should reduce catastrophic forgetting for any replay-based continual detector.","Replay selection based on decisive patches steers the model toward intrinsic manipulation cues rather than dataset-specific background, which the paper shows as higher AUC on unseen diffusion and wild datasets.","Storing fragments instead of full faces reduces identity-retrieval recall to near zero, turning replay storage into a partial de-identification mechanism.","Global lossy compression (JPEG) degrades high-frequency forgery traces, whereas spatial compression preserves original pixels in decisive areas, so the paper's spatial-selection strategy dominates quality-compression at matched storage.","The fusion scheme's combination of inter-task and intra-task mixing prevents fusion-boundary shortcuts, as evidenced by the ablation that drops performance when only inter-task fusion is used."],"fun_headline_variants":["Store only forgery hotspots to fight forgetting","Deepfake detectors: keep artifact regions, not full faces","Sparse face segments curb catastrophic forgetting","Replay decisive patches, not whole faces","Memory-efficient replay targets artifact-dense zones"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The reported wins over prior methods rest on the assumption that those baselines would fare the same on the CLIP ViT-L/14 backbone under an identical replay-memory budget, which the paper does not itself re-run.","fun_headline_variants_meta":{"raw":{"variants":["Store only forgery hotspots to fight forgetting","Deepfake detectors: keep artifact regions, not full faces","Sparse face segments curb catastrophic forgetting","Replay decisive patches, not whole faces","Memory-efficient replay targets artifact-dense zones"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000203,"raw_usage":{"total_tokens":1236,"prompt_tokens":770,"completion_tokens":466,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":398}},"tokens_in":514,"tokens_out":466,"duration_ms":4461,"temperature":1.0,"reasoning_tokens":398,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T19:40:57.761885+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the leading baselines (e.g., DFIL, SUR-LID) on the same CLIP ViT-L/14 backbone with the same replay-buffer byte budget and sample count; if their average AUC reaches or exceeds InfoDense's, the regional-selection strategy is not the source of the gain. Alternatively, swap the CLIP-similarity decisive-patch scorer for random scores while keeping the same cut-and-fuse pipeline; a non-drop would show density-awareness is not doing the work.","supporting_citations":[],"review_version":1}