{"id":"36f55c30-1cc2-4414-9b26-6d894054e6ec","arxiv_id":"2411.11396","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SUR-LID selects replay samples that approximate each task's global feature distribution and trains per-task classifiers aligned by angularity, improving incremental face forgery detection.","lead":"An incremental face forgery detector stores sparse replay subsets and isolates each task's features in latent space to reduce catastrophic forgetting. The method reports large gains over existing replay-based detectors on new benchmarks for adding forgeries one task at a time.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The core assumption that SUR replay sets remain representative after backbone updates is unverified; Eq. 7 distillation anchors only replayed points, not the old manifold, and no drift metric is reported.","rationale":"I concur with the reader's weakest_assumption. It is the most load-bearing because every component that makes SUR-LID distinct (SUR, distribution re-filling, isolation loss, and incremental decision alignment) is defined relative to the previous task's latent distribution; if that distribution is not stable under E_{t+1}, the 'bricks' are misaligned and the accumulated-knowledge story fails. The empirical results are strong enough to make the paper conditionally acceptable, but the missing drift measurement is exactly the condition that would falsify the mechanism. The bias-ignoring decision alignment (Eq. 6) is a secondary theoretical gap, but it is less likely to overturn the reported AUC gains than a systematic breakdown of replay representativeness. No formal verification or independent replication is provided, so the conditional verdict stands. My recommendation is UNCHANGED: conditional acceptance pending the drift/representativeness check and ideally seed variance reporting.","tokens_in":15670,"tokens_out":5995,"duration_ms":66604,"concrete_test":"Run Protocol 1 while retaining access to the previous training sets for diagnostics only. After each incremental step t+1, compute MMD between E_{t+1}(SUR_t) and E_{t+1}(X_t^train) (with E_t(X_t^train) as a reference) and compare with the selection-time MMD under E_t. If the drift MMD grows substantially while old-task AUC remains high, the method is robust to the assumption; if old-task AUC tracks the drift, the central claim is invalid. A complementary check is to retrain with a much larger distillation weight, e.g., mu_1=5: if performance drops, the replay set is being re-anchored rather than representing the old distribution.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the SUR replay set is a uniformly sparse proxy of each previous task's global distribution even after the backbone is fine-tuned on later tasks. SUR (Sec. 3.3) selects replay images using features from the frozen E_t, but during task t+1 training the distribution re-filling (Eq. 4), the isolation loss (Eq. 5), the detection loss (Eq. 8), and the decision alignment (Eq. 6) all operate on features of those same images extracted by the updated E_{t+1}. The only mechanism preserving the old feature space is the distillation loss Ldis (Eq. 7), which regresses E_{t+1}(\\hat{X}_i) to E_t(\\hat{X}_i) for the replayed images only. That constraint does not keep the rest of the old manifold fixed; if E_{t+1} rotates or shifts the old distribution non-uniformly, the replayed points are no longer uniformly representative, and the reconstructed 'brick' from Eq. 4 is built on drifted landmarks. The paper reports MMD (Fig. 5) as evidence that SUR approximates the training distribution, but that measurement is made with E_t at selection time, not after subsequent increments. No per-task drift statistic is provided. Therefore the claim that previous distributions are isolated and accumulated rather than overridden rests on an unverified invariance assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SUR-LID, an incremental face forgery detection method built on two components: Sparse Uniform Replay (SUR), which selects replay samples that approximate the previous task's global feature distribution, and a Latent-space Incremental Detector (LID), which adds isolation and decision-alignment losses so that each new task's feature distribution is stacked without overriding previous ones. The authors introduce two incremental protocols (dataset-incremental and forgery-category-incremental) and report large AUC gains over prior replay-based IFFD methods, together with ablations, UMAP/Grad-CAM visualizations, generalization experiments, and robustness checks. The core claim is that aligned feature isolation mitigates catastrophic forgetting and accumulates diverse forgery information.","tokens_in":15869,"tokens_out":7418,"duration_ms":74612,"significance":"If the gains are reproducible, the paper would be a meaningful advance for continual deepfake detection: the proposed protocols are more realistic than the single classical protocol used in earlier work, and the reported margins on Protocol 2 are substantial. The work also has practical strengths: code is released, ablations isolate the contribution of SUR, DR, Liso, and IDA, and the authors include cross-dataset and robustness evaluations. However, the central mechanism rests on an unverified invariance assumption about replay-set representativeness after backbone updates, the decision-alignment step does not actually align full linear decision boundaries, and the main results are reported without variance, so the significance of the numerical margins is not established.","major_comments":[{"comment":"The claim that aligning decision boundaries is identical to ensuring angularity consistency of the linear parameters is incomplete. For a linear classifier C_t(f) = theta_t * f + b_t, the decision boundary is theta_t * f + b_t = 0; Eq. (6) updates only the weight direction theta_{t+1} and says nothing about the bias b_{t+1}. Aligning weight directions makes the task-specific hyperplanes parallel, not coincident, and the inference-time averaging in Eq. (10) then combines classifiers with potentially different offsets. Please add an explicit bias-alignment mechanism, or provide evidence that bias differences are negligible in practice.","section":"Section 3.4.2 / Eq. (6)"},{"comment":"SUR selects replay samples using features from the frozen encoder E_t, but during task t+1 training the replayed images are re-embedded with the updated encoder E_{t+1}, and Eqs. (4), (5), and (8) operate on these new features. The only direct constraint on the old feature space is the distillation loss L_dis in Eq. (7), which anchors the replayed points but does not constrain the rest of the old manifold. If E_{t+1} drifts or rotates the previous feature distribution non-uniformly, the replay set is no longer a uniformly sparse proxy of the old global distribution, so the distribution re-filling in Eq. (4) is built from drifted landmarks. The MMD evidence in Fig. 5 is computed at selection time with E_t, not after subsequent increments. Please report a per-increment drift statistic, such as MMD between E_{t+1}(replay) and E_t(original training features) or per-domain feature shift, to substantiate the central stacking assumption.","section":"Sections 3.3-3.4 / Eq. (7)"},{"comment":"No variance, standard deviation, or number of independent runs is reported for any AUC value. The main comparisons and ablations therefore do not establish that the reported differences are statistically significant, especially for per-cell margins that are a few points (e.g., FF++ under Protocol 1). Please report mean +/- std over at least three seeds for the main tables and, where feasible, paired significance tests for the comparisons against prior methods and for the ablation variants.","section":"Tables 1-3 and Supplementary Table 5"}],"minor_comments":[{"comment":"The iCaRL row has arithmetic inconsistencies in the Protocol 1 averages: at T2, 0.9267 and 0.9479 average to 0.9373, not 0.8363, and at T3, 0.9010, 0.7447, and 0.9135 average to 0.8531, not 0.7864. Please correct these entries or explain the reported averages.","section":"Table 1"},{"comment":"The isolation loss denominator sums only over k with y_i != y_k, whereas the standard supervised contrastive denominator includes all other samples, including positive pairs. If the omission is intentional, please state this explicitly; otherwise the formula should be corrected.","section":"Eq. (5)"},{"comment":"The implementation details list mu1, mu2, and gamma but do not report the value of the contrastive temperature tau used in Eq. (5).","section":"Section 4.1 and Eq. (5)"},{"comment":"The displayed update for theta_{t+1} is typeset in a garbled way; please rewrite it with clear normalization and interpolation steps so that the intended operation is unambiguous.","section":"Eq. (6)"},{"comment":"There is a typo in the ablation subsection title: 'Sprase Uniform Replay' should read 'Sparse Uniform Replay'.","section":"Section 4.3"},{"comment":"The Protocol 3 results for previous methods are copied from other papers under different experimental settings, as the text acknowledges; this should be labeled as not directly comparable rather than presented as a head-to-head comparison, or the baselines should be reproduced under the same protocol.","section":"Supplementary Table 4"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a useful benchmark and a plausible method, and I do not see circularity or misconduct issues. The main concerns are technical and fixable within the manuscript's scope: the decision-alignment step needs to handle classifier bias, the replay-representativeness assumption needs a drift measurement, and the main tables need variance reporting. The arithmetic error in the iCaRL row of Table 1 should also be corrected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper is worth a serious referee. It makes a genuinely integrated contribution to incremental face forgery detection: SUR selects replay samples by magnitude, angularity, and shuffle stability; LID then isolates each task's real/fake domains with distribution re-filling and a supervised contrastive loss, and aligns task classifiers by their weight directions. The two new protocols (dataset-incremental and forgery-category-incremental) fill a real gap, and the authors actually reproduce baselines under a common setting for the main tables. The ablations show each component earns its keep, and there's code plus cross-dataset and cross-backbone generalization. That's a solid package.\n\nThe soft spots are real but mostly fixable. No error bars or multiple seeds are reported anywhere. Given that the claimed gains over HDP and DFIL are large (e.g. P2 average 0.94 vs 0.81), I find them plausible, but without variance we can't tell whether the margin is meaningful on individual datasets, especially FF++ where some numbers cluster. That is the main thing I'd require before accepting.\n\nSecond, the decision alignment in Eq. 6 only aligns the direction of the linear weight, not the bias. The boundaries are therefore not literally aligned; they're parallel but shifted. Inference by averaging classifiers may mask this, but the paper's wording 'aligned decision boundaries' overstates what Eq. 6 does. This is a theoretical mismatch, not necessarily an empirical failure.\n\nThird, the stress-test concern has teeth: SUR picks replay points using the frozen E_t, but those points are re-embedded by E_{t+1} during incremental training. The distillation loss anchors only the replayed points, so the rest of the old manifold can drift. The MMD evidence in Fig. 5 is computed at selection time, not after incrementing. If the authors added a small drift measurement (e.g. MMD between old features before/after the new task), it would directly answer this. Without it, the \"stacked brick\" story is an assumption, though a plausible one given the empirical results.\n\nThe protocol 3 comparison in the supplementary is uncontrolled — baseline numbers are copied from other papers — but the paper itself flags that, and it's not the main evidence.\n\nBottom line: the method is novel and the experiments, modulo variance, support the central claim. This deserves peer review. I'd ask for seeds/variance, a bias-alignment fix or at least a discussion, and a drift check. For the reading group, it's a yes.","headline":"Solid, well-engineered IFFD paper with a new replay strategy and feature-isolation pipeline; the gains look real but need error bars and a boundary-alignment caveat.","tokens_in":16513,"tokens_out":2496,"would_cite":true,"duration_ms":68775,"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":"Incremental face forgery detectors can avoid catastrophic forgetting by isolating each task's feature distribution in latent space and aligning their decision boundaries, the paper claims.","keywords":["incremental learning","face forgery detection","deepfake detection","catastrophic forgetting","feature isolation","replay strategy","contrastive learning","latent space"],"falsifier":"Compute, for each replayed image, the distance between its feature under the frozen old encoder and its feature under the updated encoder after each increment. If that drift is large for many replay samples, the old task's 'brick' is built from stale points; a direct experiment is to compare Protocol 1 performance when replay features are recomputed on the old backbone versus the current backbone, or when replay images are re-selected each step.","tokens_in":15384,"feed_emoji":"🧱","tokens_out":3500,"duration_ms":33830,"temperature":0.7,"pith_summary":"This paper argues that incremental face forgery detectors forget old forgeries because treating every forgery as one 'fake' class lets the new task's feature distribution overwrite the previous ones. It proposes to instead keep each task's distribution as a separate, isolated cluster in latent space, aligned so that the Real/Fake boundary stays consistent across tasks. The claim is that this 'brick-by-brick' stacking preserves old forgery knowledge while accumulating new forgery types, and the reported experiments support it with higher AUC than previous replay-based methods on two newly constructed protocols. A sympathetic reader would care because it offers a concrete way to update deepfake detectors continuously without retraining on all past data.","feed_headline":"Brick-by-brick latent stacking stops face-forgery detectors from forgetting","feed_subtitle":"Replay plus contrastive isolation lifts incremental deepfake detection AUC to 0.93–0.94 on the paper's new protocols.","key_machinery":"Two components carry the argument. Sparse Uniform Replay (SUR) selects, within each magnitude segment, the most shuffle-stable feature and a feature with the lowest angular similarity, producing a uniformly sparse version of the previous global distribution. The Latent-space Incremental Detector (LID) uses these replayed points plus distribution re-filling—mixing between replay features and their centroid—to feed a supervised contrastive isolation loss, while Incremental Decision Alignment recursively rotates the new task's linear classifier weight vector toward the previous one so all per-task decision boundaries align.","core_discovery":"The central discovery is that replay-based incremental face forgery detection works better when the replay set is chosen to approximate the whole previous distribution—not just center or hard samples—and when training explicitly separates and aligns per-task distributions. The paper reports that its Sparse Uniform Replay (SUR) selection, which picks uniformly sparse, stable samples by magnitude, angularity, and shuffle-consistency, combined with a contrastive isolation loss and decision-boundary alignment, raises average AUC from 0.8752 to 0.9315 on Protocol 1 and from 0.8141 to 0.9433 on Protocol 2 compared with HDP.","pith_inferences":["A natural extension would be to apply the SUR-LID recipe to other binary forgery or anomaly detection settings where the negative class is method-specific rather than semantically homogeneous.","The method's fixed replay budget implies a testable tradeoff: as the number of tasks grows, the per-task replay budget shrinks, and at some point distribution re-filling may no longer compensate; measuring that breakpoint would bound the method's scalability.","The shuffle-stability criterion assumes forgery features are invariant to grid shuffling; forgeries that alter global layout rather than local texture could be mis-ranked by this criterion, which is a testable extension.","One could combine SUR with generative replay to refresh old distributions, directly addressing the stale-replay drift that the paper does not measure."],"forward_implications":["Per-task feature isolation makes earlier-task accuracy degrade far less when new forgery types arrive, because old clusters are not overwritten.","Because each task's real and fake clusters stay separate, the detector accumulates diverse forgery cues instead of compressing them into a single 'fake' blob.","The aligned decision boundaries let the final detector average all task classifiers at inference, so it can exploit accumulated knowledge even when the task identity is unknown.","The approach also improves generalization to unseen datasets and to other backbone architectures, indicating the accumulated distributions carry generalizable forgery information."],"supporting_citations":[{"why":"DFIL is the replay-based IFFD method whose center-and-hard replay strategy SUR is compared against and outperforms.","marker":"[30]"},{"why":"HDP is the previous replay-based IFFD method that the paper's Protocol 1 and Protocol 2 results are mainly compared with.","marker":"[39]"},{"why":"CoReD provides the distillation-based IFFD baseline and the prior use of distillation loss that LID extends.","marker":"[19]"},{"why":"DMP supplies the mixed-prototype replay baseline and the Protocol 3 comparison results.","marker":"[41]"},{"why":"Supervised contrastive learning is the basis of the isolation loss Liso that separates per-task distributions.","marker":"[18]"},{"why":"MMD is the metric used to show that SUR replay sets approximate the original training distribution better than other replay strategies.","marker":"[12]"},{"why":"The standardized benchmark and preprocessing settings are used to ensure fair comparisons across all methods.","marker":"[45]"}],"fun_headline_variants":["Isolated features curb forgetting in incremental deepfake detection","Sparse replay plus contrastive isolation beats forgetting in deepfake detection","Aligned isolation stops face forgery detectors from forgetting","Sparse replay and feature isolation cut catastrophic forgetting in deepfake ID"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the SUR replay set, selected with the frozen backbone, still resembles the old task's global feature distribution after the backbone is fine-tuned on the new task, since all isolation and alignment losses depend on those replayed points.","fun_headline_variants_meta":{"raw":{"variants":["Isolated features curb forgetting in incremental deepfake detection","Sparse replay plus contrastive isolation beats forgetting in deepfake detection","Aligned isolation stops face forgery detectors from forgetting","Sparse replay and feature isolation cut catastrophic forgetting in deepfake ID"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001056,"raw_usage":{"total_tokens":4430,"prompt_tokens":942,"completion_tokens":3488,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":558,"completion_tokens_details":{"reasoning_tokens":3419}},"tokens_in":558,"tokens_out":3488,"duration_ms":21185,"temperature":1.0,"reasoning_tokens":3419,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:32:46.621269+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute, for each replayed image, the distance between its feature under the frozen old encoder and its feature under the updated encoder after each increment. If that drift is large for many replay samples, the old task's 'brick' is built from stale points; a direct experiment is to compare Protocol 1 performance when replay features are recomputed on the old backbone versus the current backbone, or when replay images are re-selected each step.","supporting_citations":[{"cited_title":"Dfil: Deepfake incremental learning by exploiting domain- invariant forgery clues","cited_arxiv_id":null,"evidence_quote":"DFIL is the replay-based IFFD method whose center-and-hard replay strategy SUR is compared against and outperforms."},{"cited_title":"Continual face forgery detection via historical distribution preserving","cited_arxiv_id":null,"evidence_quote":"HDP is the previous replay-based IFFD method that the paper's Protocol 1 and Protocol 2 results are mainly compared with."},{"cited_title":"Cored: Gen- eralizing fake media detection with continual representation using distillation","cited_arxiv_id":null,"evidence_quote":"CoReD provides the distillation-based IFFD baseline and the prior use of distillation loss that LID extends."},{"cited_title":"Dynamic mixed-prototype model for incremental deepfake detection","cited_arxiv_id":null,"evidence_quote":"DMP supplies the mixed-prototype replay baseline and the Protocol 3 comparison results."},{"cited_title":"Supervised contrastive learning","cited_arxiv_id":null,"evidence_quote":"Supervised contrastive learning is the basis of the isolation loss Liso that separates per-task distributions."},{"cited_title":"A kernel two-sample test","cited_arxiv_id":null,"evidence_quote":"MMD is the metric used to show that SUR replay sets approximate the original training distribution better than other replay strategies."}],"review_version":1}