{"id":"026dd1c6-4456-48d2-888b-66ccb35d51fb","arxiv_id":"2608.06876","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"FedVAR aggregates per-client normality prototypes into one global anchor and re-centers CLIP visual and text features around it, improving federated fine-grained video anomaly recognition.","lead":"The paper introduces FedVAR, a federated learning method for recognizing video anomalies like fights or accidents without sharing raw footage across cameras. It aligns all cameras to a shared 'normal scene' prototype, which the paper says reduces the confusion caused by different cameras seeing different kinds of normal behavior.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The untested cross-modal subtraction in Eq. 9 is the load-bearing assumption: subtracting a visual normality prototype from text embeddings is never validated, and no ablation isolates it.","rationale":"The reader's weakest_assumption correctly locates the foundational risk in Eq. 9. FedVAR's central novelty is the global prototype, and the mechanism works by subtracting the same m_g from both visual and textual features. If this subtraction is not semantically coherent for text, the 'cross-modal alignment' claim collapses even though the empirical global-versus-local ablation may still show gains, because those gains could come from simple visual centering or from a regularization effect rather than from true cross-modal anchoring. The paper's existing ablations do not isolate the text-side subtraction, so the assumption is genuinely untested. I also considered two other concerns: the convergence Corollary 1 overreach and the XD-Violence performance gap. Corollary 1 indeed misreads an upper bound on Gamma_local as evidence that the unaligned baseline is never better, which is not logically implied; and the abstract's 'consistently outperforms' is contradicted by Tables 4 and 5. However, both are fixable with wording changes or a corrected proof sketch, whereas a failure of Eq. 9 would undermine the core mechanism itself. The proposed ablation and cosine-alignment check settle the concern directly and are inexpensive to run. Therefore the verdict remains CONDITIONAL: the method is plausible and well-supported empirically in most settings, but the central cross-modal assumption must be tested before the mechanism can be accepted as claimed.","tokens_in":26675,"tokens_out":11710,"duration_ms":131975,"concrete_test":"Run the federated pipeline of Section 6.4 on UCF-Crime under the Random split with a single change: set d_c = E_text([t_ctx,t_c]) without subtracting m_g, while keeping Eq. 8 unchanged. Compare the resulting mAUC against the 91.06 reported in Table 2. If this variant matches or exceeds FedVAR, the cross-modal subtraction is not essential and the claimed alignment mechanism is unsupported; if it drops substantially, the subtraction matters. As a second check, on the trained FedVAR evaluate held-out cosine similarity between d_c and (x_a - m_g) for anomalous frames of each class versus (x_n - m_g) for normal frames; if this gap is not consistently positive across classes, the geometric premise of Eq. 9 fails even if prompt learning partially masks it.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is Eq. 9: d_c = E_text([t_ctx,t_c]) - m_g. FedVAR subtracts a visual normality prototype m_g, an average of CLIP image embeddings of normal frames, from text embeddings, and then treats the result as a shared anomaly direction in the joint embedding space. This assumes CLIP's joint embedding is translation-invariant in the direction of m_g, so that the same visual offset is a valid semantic origin in text space and that subtracting it yields a meaningful anomaly direction. The paper provides no direct test of this assumption. Table 9 ablates global versus local prototypes, but both arms subtract the chosen prototype from the text embedding, so the cross-modal cancellation itself is never isolated. Because the global prototype is broadcast before training and shared by all clients, any error in the text-space subtraction is a systematic bias that the learnable context tokens must compensate for; with a frozen text encoder and a single shared context vector, nothing guarantees such a compensation exists. If Eq. 9 is not semantically coherent, the shared anchor aligns only visual features, and the claimed cross-modal alignment of anomaly direction vectors loses its foundation. The same subtraction appears in the centralized AnomalyCLIP baseline, but that does not reduce its load-bearing status in FedVAR; it is inherited without independent validation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FedVAR, a federated learning framework for weakly supervised fine-grained video anomaly recognition (VAR). Each client computes a local normality prototype from frozen CLIP image embeddings; the server aggregates these into a global prototype m_g, which is broadcast to all clients and used to re-center both visual frame features and text prompt embeddings before local training of a class-specific prompt context and an axial-temporal transformer. The central claim is that this shared semantic anchor mitigates semantic misalignment under non-IID data and yields state-of-the-art results across UCF-Crime, XD-Violence, and ShanghaiTech under random, event, and scene splits, together with cross-domain and unseen-class generalization and a convergence analysis.","tokens_in":26936,"tokens_out":5793,"duration_ms":58983,"significance":"The problem is well motivated and the proposed mechanism is simple, interpretable, and aligned with the stated privacy constraints. The main empirical support is not circular: Table 9 directly compares global against local prototypes and shows consistent gains under all three partitioning schemes. The evaluation is broad, covering non-IID splits, cross-domain transfer, unseen classes, partial participation, label noise, and communication overhead, and the paper is transparent about several limitations. If the cross-modal cancellation in Eq. (9) is validated, the framework would be a useful new baseline for federated multi-class anomaly recognition. However, the current \"consistently outperforms\" claim and the comparative convergence guarantee outrun the evidence.","major_comments":[{"comment":"The load-bearing assumption that subtracting the visual normality prototype m_g from a text embedding yields a meaningful anomaly direction is never directly tested. Table 9 ablates global versus local prototypes, but both arms subtract the chosen prototype, so the cross-modal cancellation itself is not isolated. Because m_g is computed once and broadcast before training, any systematic error in this text-space subtraction is a shared bias that a frozen text encoder and a single learnable context vector must compensate for, with no guarantee that such compensation exists. I recommend a targeted analysis: compare d_c = E_text([t_ctx, t_c]) - m_g against directions derived from actual anomalous visual features, and include a control ablation that replaces m_g with a random or shifted vector of the same dimension.","section":"Section 4.2, Eq. (9)"},{"comment":"The headline claim that FedVAR \"consistently outperforms\" state-of-the-art federated baselines is contradicted by the paper's own tables. Fed-WSV AD achieves higher mAP on XD-Violence in Table 4 (49.33 vs. 47.93), higher AP on XD-Violence in Table 5 (77.33 vs. 75.20), higher cross-domain AP in Table 7 (65.42 vs. 60.20), and higher unseen-class AP in Table 8 (77.69 vs. 73.08). Section 8 acknowledges a gap on XD-Violence recognition, but the cross-domain and unseen-class gaps are not acknowledged there. The abstract and conclusion should be revised to state that FedVAR leads on UCF-Crime and ShanghaiTech while remaining competitive on XD-Violence.","section":"Abstract, Section 7.1, Tables 4, 5, 7, 8"},{"comment":"The comparative convergence claim is not supported by the stated inequalities. Lemma 1 (Eq. 17) gives an upper bound on the unaligned divergence Gamma^2_local, and Proposition 1 assumes a separate pointwise heterogeneity bound at local prototypes; neither result establishes that the actual Gamma^2_local is larger than Gamma^2_global. Corollary 1 then concludes that the unaligned baseline's error floor is \"never smaller\" and that more communication rounds are required as Delta_m grows. Comparing upper bounds does not compare actual convergence rates, and a method with a looser bound can perform better. The theory section should be reframed as two independent bounds under different assumptions, and any comparative statement should be derived from matching lower bounds or explicitly added assumptions.","section":"Section 5, Lemma 1 and Corollary 1"}],"minor_comments":[{"comment":"The method name is rendered inconsistently as \"FedV AR\", \"FedVAR\", and \"FedV AR\"; please standardize the spelling in the title, abstract, and body.","section":"Throughout"},{"comment":"The text says the results cover all three partitioning strategies, but Table 5 reports only UCF-Crime and XD-Violence; ShanghaiTech is missing from the table, so the scope should be stated explicitly.","section":"Section 7.2, Table 5"},{"comment":"Equation (20) uses E both for the number of local epochs and for the expectation operator; adopting a distinct symbol such as E_loc would remove the ambiguity.","section":"Section 5, Eq. (20)"},{"comment":"The proof sketches defer the main induction to references [43, 44]; since Theorem 1 and Proposition 1 are structurally identical to FedAvg with m_g or m_i held fixed, the paper should reproduce the key induction or clearly state which steps are new.","section":"Section 5, Proof Sketches"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope and the empirical setup is careful, but the abstract and conclusion overstate consistency, and the convergence comparison in Corollary 1 is not valid as stated. The most important fix is a focused experiment that validates or isolates the cross-modal subtraction in Eq. (9); without that, the central mechanism rests on an untested linearity assumption."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know: this is the first federated framework for fine-grained video anomaly recognition, and the central ablation supports the claimed mechanism. The headline \"consistently outperforms\" is too strong—their own Table 4 shows they trail Fed-WSV AD on XD-Violence (47.93 vs 49.33 mAP)—but they say so plainly in Section 8. The real problems are a convergence corollary that overreaches and an untested cross-modal subtraction assumption in Eq. 9.\n\nWhat's new: the problem formulation and a clean adaptation of AnomalyCLIP's normality prototype to FL. Instead of each client using its own local prototype, they aggregate a sample-weighted global prototype and freeze it before training. Table 9 shows global beats local on every split (UCF and XD, all partitioning schemes), which is real, non-circular evidence for the core mechanism. Experiments cover three benchmarks, random/event/scene splits, cross-domain transfer, and unseen-class generalization. The numbers are internally consistent, and the efficiency table (params, GFLOPs, communication) is useful. They also candidly discuss the XD-Violence gap and the risk of prototype degradation under extreme noise.\n\nSoft spots: Table 9 never isolates the cross-modal cancellation: both arms subtract the chosen prototype from text embeddings, so the paper never tests whether d_c = E_text([t_ctx, t_c]) - m_g is semantically coherent. That assumption is inherited from AnomalyCLIP, but here it is load-bearing: the global prototype is broadcast to all clients, so any systematic error in the text-space subtraction gets baked into every client's aligned space. A direct test—e.g., comparing against a text-only anchor or an additive prompt bias—would settle it. Second, Corollary 1 claims the unaligned baseline needs strictly more communication rounds whenever the misalignment term dominates, but Lemma 1's bound is an upper bound, not a lower bound. You cannot conclude a worse error floor or strictly more rounds from an upper bound; the unaligned method could in principle converge faster. That claim should be reworded as a heuristic or dropped. Minor: no code, no seeds, no standard deviations despite \"three runs,\" and the convergence proof sketches lean heavily on standard references—acceptable but not a standalone theoretical contribution.\n\nWho this is for: researchers working on federated video understanding or anomaly recognition. It is a solid engineering contribution with a well-designed ablation and an honest limitations section. It deserves a serious referee; the issues are addressable in revision.\n\nRecommendation: engage with it. Send it to peer review, and ask the authors to (1) provide a sanity check isolating Eq. 9's cross-modal subtraction, and (2) fix or soften Corollary 1.","headline":"A legitimate first federated fine-grained VAR framework with a clean mechanism and honest limitations, but the convergence analysis overreaches in Corollary 1 and the cross-modal subtraction in Eq. 9 is never directly tested.","tokens_in":27494,"tokens_out":2283,"would_cite":true,"duration_ms":24966,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T45","68T05","68W15"],"pacs":[],"model":"deepseek-v4-flash","headline":"FedVAR claims a single global normality prototype resolves semantic misalignment in federated multi-class video anomaly recognition, outperforming state-of-the-art baselines.","keywords":["federated learning","video anomaly recognition","weakly supervised learning","prototype alignment","vision-language models","prompt learning","non-IID data","semantic misalignment"],"falsifier":"Take a held-out client and compare, for frames of a given anomaly class c, the alignment between re-centered visual deviations x'_t = E_image(I_t) - m_g and the text direction d_c = E_text([t_ctx, t_c]) - m_g. If the average cosine similarity of true anomalous frames to d_c is not significantly above that of normal frames, or if a zero-shot ranking built from d_c without any training performs at chance, the cross-modal subtraction that the whole alignment rests on is not empirically meaningful.","tokens_in":1559,"feed_emoji":"🎥","tokens_out":4370,"duration_ms":73002,"temperature":0.7,"pith_summary":"FedVAR is a federated framework for fine-grained video anomaly recognition. It identifies semantic misalignment, the divergence of client feature spaces under non-IID data, as the main barrier to multi-class anomaly recognition. Its proposal is to aggregate client-side normality prototypes into one global prototype and re-center every client's visual and textual CLIP features around that shared anchor. If correct, this makes weakly supervised multi-class anomaly recognition feasible in privacy-preserving edge networks without exchanging raw video. The paper reports consistent gains over state-of-the-art federated baselines on UCF-Crime, XD-Violence, and ShanghaiTech across random, event, and scene splits, and it shows transfer to unseen domains and unseen anomaly classes.","feed_headline":"One global normality anchor lifts federated video anomaly recognition","feed_subtitle":"FedVAR re-centers every client's features on one shared prototype, beating federated baselines on three benchmarks.","key_machinery":"The global normality prototype $m_g = \\frac{\\sum_i |N_i| m_i}{\\sum_i |N_i|}$ is a weighted average of client-side visual prototypes of normal frames. It is broadcast to all clients, and every visual frame feature and every textual anomaly-class prompt is then re-centered by subtracting $m_g$, defining anomaly direction vectors $d_c = E_{\\mathrm{text}}([t_{\\mathrm{ctx}}, t_c]) - m_g$ that are comparable across clients. Anomaly likelihood is the projection of a re-centered frame onto $d_c$, combined with an Axial Transformer temporal module and a weakly supervised multiple-instance-learning objective.","core_discovery":"The paper claims to establish that semantic misalignment in federated video anomaly recognition can be resolved by a single shared normality anchor. Each client computes a local prototype from its own normal frames, the server aggregates these into a global prototype, and all clients then re-center both visual frame features and text-prompt embeddings around that anchor. This makes anomaly direction vectors consistent across clients with disjoint anomaly classes and heterogeneous scenes. Empirically, on UCF-Crime, XD-Violence, and ShanghaiTech, under random, event, and scene splits, the method reports state-of-the-art mean AUC and mean AP among federated baselines, and it demonstrates generalization to unseen domains and unseen anomaly classes.","pith_inferences":["The same prototype re-centering recipe may transfer to other federated tasks with a well-defined \"normality\" concept, such as industrial defect detection or sensor-based anomaly monitoring, whenever a shared embedding space exists.","A direct test of the linearity assumption would be to compare the global-prototype anchor against a text-only prototype or a random vector; the size of the performance gap would reveal how much of the gain comes from the semantic content of the anchor rather than from simple recentering.","The paper's passing remark that the prototype can be updated online with an exponential moving average suggests a natural formal extension for handling client churn and long-term distribution shift, which is otherwise left implicit.","The slight gap to Fed-WSVAD on XD-Violence suggests that combining prototype alignment with finer temporal localization is a promising extension, a direction the paper itself identifies."],"forward_implications":["Federated multi-class anomaly recognition becomes achievable with only video-level labels and no raw data exchange between clients.","A shared normality anchor provides a transferable representation, yielding improved cross-domain and unseen-class generalization relative to methods that only aggregate prompts or parameters.","Communication overhead stays small because the prototype is one D-dimensional vector uploaded once per client, and the trainable module is limited to prompt tokens and a lightweight temporal model.","The convergence analysis implies that unaligned local prototypes carry an extra gradient-divergence penalty proportional to prototype spread, so aligned training reaches a target accuracy in no more (and often fewer) communication rounds than the unaligned baseline.","The framework stays stable under partial client participation and moderate label noise, since the weighted prototype aggregation filters sparse corruption."],"supporting_citations":[{"why":"Supplies the frozen CLIP visual-text embedding space that FedVAR re-centers and shares across clients.","marker":"[12]"},{"why":"AnomalyCLIP is the centralized architecture whose normality-prototype recentering FedVAR extends to the federated setting.","marker":"[13]"},{"why":"FedAvg is the parameter aggregation protocol used to combine client model updates.","marker":"[16]"},{"why":"Fed-WSVAD is the main federated weakly-supervised video anomaly detection baseline that FedVAR must beat.","marker":"[17]"},{"why":"UCF-Crime provides the weak video-level labels, the MIL formulation, and the largest benchmark used in the evaluation.","marker":"[10]"},{"why":"XD-Violence provides the long-video benchmark and the average-precision evaluation protocol.","marker":"[45]"},{"why":"ShanghaiTech provides the scene-diverse benchmark used to test scene-dependent normality.","marker":"[46]"},{"why":"CLAP is the federated unsupervised anomaly detection baseline compared against in the experiments.","marker":"[18]"}],"fun_headline_variants":["One normality anchor fixes federated video anomaly drift","Single shared prototype aligns all federated video anomaly clients","FedVAR stops semantic drift in federated video anomaly recognition","One prototype re-centers all federated video anomaly clients","FedVAR's single anchor aligns video anomaly views across clients"],"cache_read_input_tokens":29568,"weakest_assumption_plain":"The method assumes that a normal visual prototype computed from normal frames can be subtracted from text embeddings of anomaly class names to produce a meaningful anomaly direction, meaning CLIP's visual and text spaces are additively compatible; if that linearity breaks, the global anchor cannot align clients.","fun_headline_variants_meta":{"raw":{"variants":["One normality anchor fixes federated video anomaly drift","Single shared prototype aligns all federated video anomaly clients","FedVAR stops semantic drift in federated video anomaly recognition","One prototype re-centers all federated video anomaly clients","FedVAR's single anchor aligns video anomaly views across clients"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001724,"raw_usage":{"total_tokens":6825,"prompt_tokens":958,"completion_tokens":5867,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":5789}},"tokens_in":574,"tokens_out":5867,"duration_ms":41293,"temperature":1.0,"reasoning_tokens":5789,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T19:09:19.293798+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out client and compare, for frames of a given anomaly class c, the alignment between re-centered visual deviations x'_t = E_image(I_t) - m_g and the text direction d_c = E_text([t_ctx, t_c]) - m_g. If the average cosine similarity of true anomalous frames to d_c is not significantly above that of normal frames, or if a zero-shot ranking built from d_c without any training performs at chance, the cross-modal subtraction that the whole alignment rests on is not empirically meaningful.","supporting_citations":[{"cited_title":"Radford, J","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen CLIP visual-text embedding space that FedVAR re-centers and shares across clients."},{"cited_title":"Zanella, B","cited_arxiv_id":null,"evidence_quote":"AnomalyCLIP is the centralized architecture whose normality-prototype recentering FedVAR extends to the federated setting."},{"cited_title":"McMahan, E","cited_arxiv_id":null,"evidence_quote":"FedAvg is the parameter aggregation protocol used to combine client model updates."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Fed-WSVAD is the main federated weakly-supervised video anomaly detection baseline that FedVAR must beat."},{"cited_title":"Sultani, C","cited_arxiv_id":null,"evidence_quote":"UCF-Crime provides the weak video-level labels, the MIL formulation, and the largest benchmark used in the evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"XD-Violence provides the long-video benchmark and the average-precision evaluation protocol."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ShanghaiTech provides the scene-diverse benchmark used to test scene-dependent normality."},{"cited_title":"Al-Lahham, M","cited_arxiv_id":null,"evidence_quote":"CLAP is the federated unsupervised anomaly detection baseline compared against in the experiments."}],"review_version":1}