{"id":"6c6921e7-b78a-4f43-996d-2f39273672ad","arxiv_id":"2412.03044","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A frequency-guided diffusion model with perturbation training improves skeleton-based video anomaly detection accuracy on five benchmarks.","lead":"This paper trains a diffusion model for skeleton-based video anomaly detection with adversarial perturbation samples that make the model robust to unseen normal motions, plus a DCT-based frequency guide that copies local motion details from the observed input while reconstructing global structure. The method reports improved accuracy over prior work on five public benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The frequency premise in Eq. (18) is the load-bearing risk: anomalies expressed only in high-frequency DCT coefficients are copied into the reconstruction and never scored; the paper never tests this failure mode.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing concern: the method assumes anomalies are low-frequency global structure while high-frequency content is person-specific style, and Eq. (18) copies observed high-frequency coefficients into the reconstruction. This is the most direct threat to the central claim because it predicts a concrete class of anomalies that FG-Diff is structurally unable to detect. The flawed Theorem IV.1 is a genuine correctness defect, but it does not by itself invalidate the empirical SOTA claim; it can be removed or corrected without altering the method. The high-frequency-copy failure mode, however, is intrinsic to the inference procedure. The paper does provide useful empirical support: ablations show DCT-Mask and perturbation training each improve AUC, and the robustness analysis in Table IV shows adversarial training helps. But those results do not test the frequency premise. A synthetic high-frequency anomaly experiment would settle whether the premise is valid. If the test shows FG-Diff misses high-frequency anomalies, the claim 'more accurate and robust anomaly detection' should be restricted to low-frequency-dominant anomalies. If the test shows FG-Diff still detects them (e.g., because the final score is dominated by low-frequency reconstruction errors or because the perturbation generator redistributes the signal), the concern is resolved. Given the reader already recommended CONDITIONAL acceptance, my analysis does not move the verdict; it sharpens the condition that should be imposed.","tokens_in":17894,"tokens_out":4540,"duration_ms":49843,"concrete_test":"Construct a high-frequency-only anomaly test set from normal skeleton sequences: take normal motions x, compute y = DCT(x), perturb or zero only the coefficients outside the top λ_dct low-frequency set (e.g., coefficients above a temporal-frequency cutoff), and reconstruct x_a = IDCT(y_a). These sequences retain near-identical low-frequency global structure while carrying abnormal high-frequency detail. Run FG-Diff with its published λ_dct values and a baseline method such as MoCoDAD or TrajREC on the normal and high-frequency-anomalous sequences, and compute per-sequence AUC. If FG-Diff's AUC is near chance while the baseline detects these anomalies, then Eq. (18)'s copying of observed high-frequency coefficients is confirmed as a blind spot and the central robustness claim is overgeneralized.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that FG-Diff is a more accurate and robust skeleton VAD detector rests on the premise that high-frequency DCT coefficients encode person-specific style rather than anomaly-relevant structure. This premise is operationalized in Eq. (18): y_c^t = y_o^t ⊙ M_h(y_o^t) + y_g^t ⊙ M_l(y_g^t). At every denoising step, the observed motion's high-frequency coefficients are copied unmodified into the fused motion, and only low-frequency coefficients are regenerated. If an anomalous event manifests primarily as a high-frequency deviation -- a brief local hand movement, a subtle gesture, or a temporally localized joint spike -- those anomalous coefficients are inserted into the reconstruction, so the final reconstruction error S(x) = ||x − x_g||_2^2 remains small. The anomaly is then missed by construction. The paper's defense of this design is the qualitative claim that high-frequency information is person-specific style, but no experiment isolates this case. The threshold λ_dct is also tuned per dataset (0.1 for Avenue/HR-Avenue/HR-STC, 0.9 for UBnormal/HR-UBnormal), so the split between 'global' and 'detail' is not a stable semantic property; Fig. 6 reports sensitivity only for HR-Avenue and HR-STC. A secondary defect is Theorem IV.1: the proof of increased reconstruction error, Eq. (26), is invalid because maximization of the training loss over φ does not imply S(x̂_o) ≥ S(x_o), and the '⇐⇒' step is unjustified. This theorem is correctable, but the frequency-copy mechanism directly threatens the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FG-Diff, a skeleton-based video anomaly detection method combining a diffusion model trained with adversarial perturbation samples (a learned perturbation generator) and a frequency-guided denoising process. At training time, a perturbation generator maximizes the denoising loss within a bounded neighborhood, and the denoiser minimizes it on the perturbed samples. At inference, the method applies 2D-DCT masks to fuse high-frequency components of the observed motion with low-frequency components of the generated motion at each denoising step, then scores anomalies by reconstruction error. The authors report state-of-the-art AUC on Avenue, HR-Avenue, HR-STC, UBnormal, and HR-UBnormal, together with ablations and a robustness analysis.","tokens_in":18224,"tokens_out":4172,"duration_ms":38157,"significance":"If the empirical results hold, the paper would contribute a practical recipe for improving open-set robustness of skeleton-based reconstruction methods, and the frequency-separation idea is an interesting departure from uniform reconstruction targets. The paper is commendable for evaluating on five datasets, including open-set UR-normal, for providing a robustness analysis of perturbation training, and for reporting parameter efficiency against supervised baselines. However, the central theoretical justification (Theorem IV.1) is invalid, and the frequency-copy fusion rule has an untested failure mode that is load-bearing for the anomaly-detection claim. The per-dataset tuning of the frequency threshold and the lack of statistical significance reporting further weaken the strength of the empirical claims. The core idea is defensible, but the manuscript needs substantial revision before the claims can be accepted.","major_comments":[{"comment":"The above comment is a major comment with invalid proof. Please provide a corrected version.","section":"Section VII-A / Theorem IV.1 (Eqs. 25-27)"},{"comment":"The above comment is a major comment. Please provide a corrected version.","section":"Section IV-C, Eq. (18), and Algorithm 2"},{"comment":"The above comment is a major comment. Please provide a corrected version.","section":"Algorithm 2, lines 6-7, and Table IV"},{"comment":"The above comment is a major comment. Please provide a corrected version.","section":"Tables I-III"}],"minor_comments":[{"comment":"The above comment is a minor comment. Please provide a corrected version.","section":"Eq. (22)"},{"comment":"The above comment is a minor comment. Please provide a corrected version.","section":"Algorithm 2, line 6"},{"comment":"The above comment is a minor comment. Please provide a corrected version.","section":"Section V-C, Table III"},{"comment":"The above comment is a minor comment. Please provide a corrected version.","section":"Fig. 6"}],"recommendation":"major_revision","confidential_remarks":"The invalid proof of Theorem IV.1 is likely to be flagged by reviewers; the authors should either correct it or remove the theorem. The frequency-copy failure mode is the most serious scientific risk and needs a dedicated experiment. Also, the per-dataset threshold tuning reduces confidence in the method's generality; consider reporting sensitivity on more datasets and, ideally, releasing code to verify the results. The paper otherwise fits the journal's scope and the empirical direction is promising."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you work on skeleton-based VAD. The paper combines two ideas that are individually known—adversarial perturbation training and frequency-domain separation—but the combination is new to this task, and the empirical work is consistent. On five benchmarks the reported AUC gains over strong baselines are modest (0.6–3.1 points), but they hold across all five, and the ablations back up both components. The robustness analysis in Table IV is particularly nice: training with the perturbation generator keeps AUC stable when you increase perturbation at inference, while the model without it degrades sharply.\n\nThe soft spots are real but mostly fixable. Theorem IV.1's proof is wrong. The step that turns 'the generator maximizes the diffuser's training loss' into 'the perturbed motion is the argmax of reconstruction error' is invalid; the biconditional in Eq. (26) does not follow. Since the empirical claims don't depend on the theorem, the right move is to delete or correct it. More substantive: the DCT fusion rule in Eq. (18) copies observed high-frequency coefficients straight into the reconstruction. The authors assert high-frequency content is person-specific style, but if an anomaly lives in a local high-frequency detail—a brief hand movement, a joint spike—those anomalous coefficients are preserved and the anomaly score stays low. The paper never tests that failure mode. The per-dataset tuning of λ_dct (0.1 vs 0.9) suggests the 'global vs. detail' split is not a stable semantic property, and the sensitivity curves only cover two datasets. Also, the inference-time perturbation injection into both observed and generated motions is conceptually odd; the ablation shows it doesn't hurt, but the paper doesn't explain why it helps.\n\nStill, the central empirical claim is not circular and the method is clearly better than the previous SOTA on the reported benchmarks. No error bars or significance tests, and no code released, so the gains should be treated as preliminary. I'd send this to peer review: the idea is worth developing, and a good referee can push for the high-frequency anomaly test and a cleaned-up theorem. If you're in the area, it's a useful related work.","headline":"Solid empirical method paper for skeleton-based VAD with a useful perturbation-training + DCT-guidance combination, but the proof of Theorem IV.1 is invalid and the frequency-copy mechanism may miss high-frequency anomalies.","tokens_in":18773,"tokens_out":3530,"would_cite":true,"duration_ms":33605,"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":"A frequency-guided diffusion model with perturbation training reports state-of-the-art accuracy on five skeleton-based video anomaly detection benchmarks by reconstructing global motion while preserving observed high-frequency detail.","keywords":["video anomaly detection","skeleton-based detection","diffusion model","perturbation training","discrete cosine transform","frequency-guided denoising","open-set robustness","reconstruction-based detection"],"falsifier":"Build or collect skeleton sequences where the global motion is a normal walk but a single joint (for example, the wrist) performs a small rapid tremor that is abnormal in context. If the proposed method's reconstruction error for these sequences stays as low as for genuinely normal sequences, the frequency-fusion assumption fails; a simpler check is to run the released code on such sequences and compare AUC against a baseline that uses full reconstruction error.","tokens_in":17696,"feed_emoji":"🦴","tokens_out":4231,"duration_ms":37736,"temperature":0.7,"pith_summary":"This paper tries to show that two weaknesses of reconstruction-based video anomaly detection, poor robustness to unseen normal motions and unreliable reconstruction of local motion details, can both be fixed inside a single diffusion framework. The first fix is adversarial: a small generator learns perturbations that look like normal motions but maximize the diffusion model's reconstruction error, and the denoiser is trained against them to widen the set of motions it can reconstruct. The second fix is spectral: a 2D discrete cosine transform separates each motion into low-frequency (global) and high-frequency (local) parts, and at inference the observed high-frequency coefficients are copied straight into the reconstruction while only the low-frequency part is regenerated. The paper reports that this combination exceeds previous state-of-the-art AUC on Avenue, HR-Avenue, HR-STC, UBnormal, and HR-UBnormal. A sympathetic reader would care because it offers a concrete recipe for making open-set anomaly detection more forgiving of normal variation without needing anomalous training data.","feed_headline":"Frequency-guided diffusion tops five skeleton anomaly tests","feed_subtitle":"Adversarial perturbation training plus DCT-based fusion lifts open-set AUC on Avenue, STC, and UBnormal benchmarks.","key_machinery":"The load-bearing mechanism is the frequency-guided denoising loop, in which a 2D Discrete Cosine Transform (DCT) maps each motion into coefficients ordered by frequency, and a DCT-Mask splits them into a low-frequency set (the largest-magnitude coefficients, thresholded by $\\lambda_{dct}$) and the complementary high-frequency set. At every denoising step the observed motion's high-frequency coefficients are copied into the fused motion while the denoiser's low-frequency coefficients are kept; the inverse DCT then returns a motion that the denoiser refines toward the normal-motion distribution. The second mechanism is the perturbation generator $G_\\phi$, a lightweight network producing sign-constrained perturbations $\\lambda_p \\cdot \\mathrm{sign}(G_\\phi(x))$ that maximize the denoising loss, trained adversarially against the noise predictor $\\epsilon_\\theta$ so that the reconstruction domain expands toward unseen normal motions. A theorem in the appendix asserts that the generated perturbed motion stays within a bounded neighborhood of the observed motion while increasing reconstruction error.","core_discovery":"The central discovery claimed is that anomaly detection accuracy improves when a diffusion model is deliberately trained on adversarially chosen near-normal perturbations and then, at inference, is allowed to keep the high-frequency DCT coefficients of the observed motion instead of reconstructing them. The paper argues that high-frequency motion components (stride length, arm-swing amplitude, joint-level habits) vary from person to person and are therefore not reliable signals of abnormality, while low-frequency components encode the global structure that determines whether a motion is normal or anomalous. Under the fusion rule $y^c_t = y^o_t \\odot M_h(y^o_t) + y^g_t \\odot M_l(y^g_t)$, the model regenerates only low-frequency content and preserves observed high-frequency detail, which the paper shows yields higher AUC than reconstructing all frequencies equally. Combined with perturbation training, the method reaches 88.0 AUC on Avenue, 90.7 on HR-Avenue, 78.6 on HR-STC, 68.9 on UBnormal, and 69.0 on HR-UBnormal, outperforming the listed state-of-the-art methods including prediction-based and hybrid ones. The paper also shows that perturbation training keeps performance stable when inference-time perturbation intensity increases, whereas a model trained without it degrades sharply.","pith_inferences":["The frequency-splitting idea is not tied to skeletons; the same DCT-fusion rule could be tested on RGB video features or general time-series anomaly detection, where global trend versus local detail separation may behave differently.","The paper's assumption that anomalies live in low-frequency structure is testable by constructing anomalies confined to a single joint or small joint group; if such anomalies pass through with low error, the fusion rule needs an exception mechanism.","The threshold $\\lambda_{dct}$ is tuned per dataset, suggesting the optimal low/high split depends on the motion distribution; an adaptive threshold derived from training-set statistics could remove this per-dataset tuning and is a direct extension of the paper's own sensitivity analysis."],"forward_implications":["Unseen normal motions with stylistic variations should no longer be systematically misclassified, because perturbation training deliberately exposes the model to near-normal inputs it would otherwise reconstruct poorly.","Reconstruction error becomes a more reliable anomaly score, because the score no longer depends on the model's ability to reproduce person-specific high-frequency details.","The approach remains competitive with supervised and weakly supervised methods (for example, 68.9 AUC on UBnormal) while using only 556K parameters and no anomaly labels.","Ablations indicate that both components contribute: removing perturbation training drops AUC by up to 4.6 points, and replacing the DCT-Mask with a temporal mask drops it by up to 0.9 points."],"supporting_citations":[{"why":"Supplies the diffusion-based skeleton VAD baseline (MoCoDAD) that the proposed method builds on and compares against.","marker":"[8]"},{"why":"Provides the previous state-of-the-art trajectory reconstruction and prediction method (TrajREC) that the proposed method outperforms on four benchmarks.","marker":"[17]"},{"why":"Introduces the HR-Avenue and HR-STC skeleton VAD datasets and the MPED-RNN reconstruction/prediction baselines.","marker":"[7]"},{"why":"Provides the FGSM-style sign perturbation that the perturbation generator is inspired by and that the bound argument relies on.","marker":"[45]"},{"why":"Gives the DCT coefficient analysis used to justify separating motion into low- and high-frequency components.","marker":"[48]"},{"why":"Supplies the Avenue benchmark and its skeleton extraction setup used in evaluation.","marker":"[49]"},{"why":"Introduces the UBnormal open-set benchmark and its anomaly split protocol used in the experiments.","marker":"[6]"},{"why":"Establishes the AUC evaluation protocol for video anomaly detection that the paper follows.","marker":"[30]"}],"fun_headline_variants":["FG-Diff: frequency-guided diffusion wins five anomaly tests","Perturbation-trained diffusion beats skeleton anomaly SOTA","DCT split + perturbation training lifts anomaly detection","Diffusion keeps high-freq detail, detects anomalies better"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that abnormal motions always change the global, low-frequency structure of a skeleton sequence, while high-frequency components are only person-specific style, so an anomaly that lives entirely in small, local joint details would be copied into the reconstruction and missed.","fun_headline_variants_meta":{"raw":{"variants":["FG-Diff: frequency-guided diffusion wins five anomaly tests","Perturbation-trained diffusion beats skeleton anomaly SOTA","DCT split + perturbation training lifts anomaly detection","Diffusion keeps high-freq detail, detects anomalies better"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000258,"raw_usage":{"total_tokens":1640,"prompt_tokens":1060,"completion_tokens":580,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":676,"completion_tokens_details":{"reasoning_tokens":515}},"tokens_in":676,"tokens_out":580,"duration_ms":5915,"temperature":1.0,"reasoning_tokens":515,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:49:38.836350+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build or collect skeleton sequences where the global motion is a normal walk but a single joint (for example, the wrist) performs a small rapid tremor that is abnormal in context. If the proposed method's reconstruction error for these sequences stays as low as for genuinely normal sequences, the frequency-fusion assumption fails; a simpler check is to run the released code on such sequences and compare AUC against a baseline that uses full reconstruction error.","supporting_citations":[{"cited_title":"Multimodal motion conditioned diffusion model for skeleton-based video anomaly detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the diffusion-based skeleton VAD baseline (MoCoDAD) that the proposed method builds on and compares against."},{"cited_title":"Holistic representation learning for multitask trajectory anomaly detection,","cited_arxiv_id":null,"evidence_quote":"Provides the previous state-of-the-art trajectory reconstruction and prediction method (TrajREC) that the proposed method outperforms on four benchmarks."},{"cited_title":"Learning regularity in skeleton trajectories for anomaly detection in videos,","cited_arxiv_id":null,"evidence_quote":"Introduces the HR-Avenue and HR-STC skeleton VAD datasets and the MPED-RNN reconstruction/prediction baselines."},{"cited_title":"A mathematical analysis of the dct coefficient distributions for images,","cited_arxiv_id":null,"evidence_quote":"Gives the DCT coefficient analysis used to justify separating motion into low- and high-frequency components."},{"cited_title":"Abnormal event detection at 150 fps in matlab,","cited_arxiv_id":null,"evidence_quote":"Supplies the Avenue benchmark and its skeleton extraction setup used in evaluation."},{"cited_title":"Ubnormal: New benchmark for supervised open-set video anomaly detection,","cited_arxiv_id":null,"evidence_quote":"Introduces the UBnormal open-set benchmark and its anomaly split protocol used in the experiments."},{"cited_title":"Learning temporal regularity in video sequences,","cited_arxiv_id":null,"evidence_quote":"Establishes the AUC evaluation protocol for video anomaly detection that the paper follows."}],"review_version":1}