{"id":"1c6e8441-33e2-4d08-9f70-bf33de65d2e9","arxiv_id":"2507.14807","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"HICOM is a multi-face deepfake detection framework whose four modules are each inspired by cues that humans reportedly use to spot fake faces, achieving state-of-the-art frame-level complete detection on existing benchmarks.","lead":"A new deepfake detector, HICOM, uses insights from human studies to find and flag every fake face in group photos and videos, rather than just one face at a time. It reports better accuracy on multi-face benchmarks and claims to generalize better to new datasets, though several details need scrutiny.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The fusion rule in §4.6 is self-contradictory: XOR cannot implement \"any detected anomaly leads to fake\", so the reported gains are not tied to a well-defined model.","rationale":"The reader's weakest assumption (small, potentially biased human study) is real, but it is not the most load-bearing issue for the central claim. Even a perfect human study would not rescue the results if the fusion rule is not well-defined, because the architecture as written cannot produce the reported behavior under the stated XOR rule. Conversely, if the fusion rule is clarified and reproduces Tables 1 and 4, the human-study concern becomes a question of cue stability rather than a logical blocker. I therefore focus on §4.6. A supporting red flag reinforces this choice: for OpenForensics, an image-only dataset, §5.1 states that each image is replicated to form the M1 input sequence, so no genuine motion signal exists for the H1 scene-motion cue; yet Table 4 shows M1 alone contributing 93.9 FCAC there. That makes it especially important to know exactly how module outputs are combined before attributing gains to the human-inspired cues. The contradiction is resolvable by clarification and re-running the fusion ablation, so the verdict remains CONDITIONAL rather than REJECT. I do not see evidence of deliberate misreporting; the paper simply does not specify a consistent model in its current form.","tokens_in":16017,"tokens_out":6044,"duration_ms":64259,"concrete_test":"Recompute Tables 1 and 4 using both fusion rules on the same per-module outputs. Instrument the fusion stage to count faces/frames where two or more modules vote fake; for those cases XOR returns real while OR returns fake. If XOR is genuine, the M1+M2+M3+M4 row must be no better than M1+M2+M3 on at least some datasets, unless no co-positive cases exist. If OR is required to match the published numbers, §4.6 must be corrected and the central claim re-evaluated. The decisive quantities are the count of co-positive module outputs and the resulting FCAC/FCAU under each rule.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The decisive weak point is the fusion rule in §4.6. The text says HICOM fuses module outputs \"using an XOR operation, ensuring that any detected anomaly leads to a fake face prediction.\" These two specifications are mutually incompatible: XOR outputs true only when an odd number of inputs are true, so a face flagged by both M1 and M2 produces 1 XOR 1 = 0 (classified as real). That is the opposite of \"any detected anomaly.\" The surrounding rationale—\"When M1 and M2 miss a fake face, M3 and M4 help identify these inconsistencies\"—describes an OR-like rule, and Table 4's monotone progression M1 < M1+M2 < M1+M2+M3 < M1+M2+M3+M4 is only plausible under OR or another non-XOR rule. Because fusion is the only mechanism that converts the four human-inspired module outputs into the reported frame-level predictions, the 3.3% in-dataset and 5.8% cross-dataset improvements cannot be attributed to the proposed cues unless the actual fusion rule is specified and shown to reproduce the tables. The contradiction is load-bearing, not cosmetic: it changes predicted behavior on exactly the multi-cue frames the paper emphasizes.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HICOM, a multi-face deepfake detection framework whose design is motivated by a two-phase human study. The authors report four human cues (scene-motion coherence, inter-face appearance compatibility, interpersonal gaze alignment, and face-body consistency), instantiate each as a module (M1–M4), and fuse module outputs to label every face in a frame. Experiments on FFIW, OpenForensics, DF-Platter, and ManualFake report improvements in frame-level complete multi-face detection accuracy, robustness to perturbations, cross-dataset generalization, and an LLM-based explanation interface.","tokens_in":16285,"tokens_out":7875,"duration_ms":81763,"significance":"The paper's central idea—basing multi-face detection on cues elicited from human observers and evaluating a strict frame-level complete multi-face metric—is timely and potentially useful. Strengths include a clearly defined evaluation metric, a modular architecture with each module tied to a stated hypothesis, and comparisons against multiple single- and multi-face baselines. If the cue discovery and fusion specification are made rigorous and non-circular, the approach could make a solid contribution to the deepfake detection literature. At present, however, the reported gains rest on an internally contradictory fusion rule and on cue-selection evidence that is confounded with the evaluation benchmarks.","major_comments":[{"comment":"The fusion rule as written is self-contradictory. The text says the module outputs are combined 'using an XOR operation, ensuring that any detected anomaly leads to a fake face prediction,' but XOR outputs 1 only when an odd number of inputs are 1; a face flagged by exactly two modules yields 1 XOR 1 = 0, i.e., real. The stated rationale ('When M1 and M2 miss a fake face, M3 and M4 help identify these inconsistencies') and the monotone ablation in Table 4 (M1 < M1+M2 < M1+M2+M3 < M1+M2+M3+M4) both describe an OR-like rule rather than XOR. Since the fused output is the final prediction behind Tables 1–3, the reported 3.3% in-dataset and 5.8% cross-dataset gains cannot be attributed to the proposed cues unless the actual fusion rule is specified (including how each module's continuous output is binarized, any thresholds, and the exact Boolean combination) and shown to reproduce the tables. This is a load-bearing specification, not a wording issue.","section":"§4.6"},{"comment":"The cue discovery is circular with the evaluation. The four cues were elicited on the same OpenForensics, FFIW, and DF-Platter datasets used to evaluate HICOM in Tables 1 and 2; the cross-dataset test in Table 3 uses ManualFake, which was not part of cue discovery, but the design choices (which cues to include, how to threshold them) were already informed by the evaluation distribution. As a result, the 5.8% generalization gain cannot be cleanly attributed to human-inspired cues as opposed to cues that happen to be informative on these benchmarks. Please validate the cues on a held-out split or an independent deepfake type not seen during cue discovery, and at minimum explicitly discuss this confound in the limitations.","section":"§3.2, §5.3"},{"comment":"The quantitative evidence for the four cues is thin. Phase 1 recruited only four participants, each reviewing 500 samples, and §3.3 reports prevalence percentages (34.2%, 31.5%, 25.0%, 7.5%) without per-participant variance, inter-rater reliability, or significance tests. The criterion 'Only identifications matching dataset labels were considered valid' further conditions the cue reporting on label consistency, which may suppress cues that lead to false alarms. With n=4, these percentages cannot support the claim that the four cues are the key human strategies. Please report the full distribution of cue frequencies across participants, agreement metrics, and the number of samples contributing to each percentage.","section":"§3.2, §3.3"},{"comment":"The gaze module's decision rule is ad hoc and not grounded in the human study. The conditions 'n_L - n_O > 1 or n_T = 2' appear without derivation or ablation; similarly, the contrastive margin and λcomp in Eq. (2) are set empirically. Since Table 4 attributes incremental gains to M3 and M4, these hand-set thresholds could be the source of the improvement rather than the human-inspired cue itself. Please ablate these thresholds and report sensitivity, or justify them from the human-study data.","section":"§4.4, Eq. (3)"}],"minor_comments":[{"comment":"The table header contains the typo 'OpenForencics'; it should read 'OpenForensics'.","section":"Table 1"},{"comment":"The text refers to a Supplementary Material for the fusion strategy, but the submitted manuscript does not appear to include that supplement; either provide it or remove the reference.","section":"§4.6"},{"comment":"The human-comparison result is reported only as 'three stars'; please give the actual test statistic, p-value, and sample size.","section":"§5.4, Fig. 5"},{"comment":"Reference [23] is cited for the 'Dual-path Actor Interaction framework with Multi-scale Actor Contrastive Loss,' which does not match the Gallagher and Chen paper; the citation appears to be incorrect.","section":"Related Work"},{"comment":"The abstract says 'average accuracy by 3.3%,' while the body specifies 'average improvements of 3.3% in FCAC, and 3.1% in FCAU'; clarify which metric is being summarized.","section":"Abstract / §5.2"},{"comment":"Phase 2 reports 20 MTurk participants but does not state the number of judgments per item, compensation, or exclusion criteria; adding these details would improve reproducibility.","section":"§3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper's central experiment set is interesting, but the fusion-rule contradiction and the circularity of cue discovery are serious. The manuscript is salvageable with a major revision that specifies the fusion rule, re-runs or reinterprets the ablations, and strengthens the human-study evidence. I would not recommend acceptance in the current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing you should know first: the fusion rule in §4.6 is self-contradictory, and it is load-bearing. The text says HICOM fuses module outputs “using an XOR operation, ensuring that any detected anomaly leads to a fake face prediction.” XOR does not do that: 1 XOR 1 = 0, so a face flagged by both M1 and M2 would be classified real. The surrounding prose and Table 4’s monotone ablation are only consistent with an OR-like (or soft/learned) rule. This is not cosmetic — fusion is the only mechanism that connects the four module outputs to the frame-level predictions, so the reported 3.3% in-dataset and 5.8% cross-dataset gains are not tied to a well-defined model until the authors specify the actual fusion and show it reproduces the tables. The stress-test note is correct.\n\nThat said, the paper does real things well. The two-phase human study is a genuine attempt to derive cues from human perception rather than researcher intuition, and frame-level complete multi-face detection is an under-studied evaluation mode worth pushing. The four cues are plausible, the modular design is clearly described (aside from fusion), and the experiments span in-dataset, perturbed, cross-dataset, and human-comparison settings. The reported gains are sizable if they hold.\n\nBeyond the fusion issue, the soft spots are: the phase-1 human study used four participants and validated identifications against dataset labels, which risks overfitting cue discovery to the same benchmarks used for evaluation (a mild circularity). The abstract’s “2.8% under real-world perturbations” does not match the section text (1.5% FCAC, 2.8% FCAU). No code or data release, and no error bars on the core tables. These are fixable with disclosure.\n\nWho is this for? People working on multi-face or context-aware deepfake detection, and anyone designing human-study-driven pipelines. It deserves a serious referee: the core idea is valuable, the main flaw is correctable, and the authors should be asked to fix the fusion description, correct the numbers, report variance, and release code. I would not cite it in its current form, but I would send it to review.","headline":"A load-bearing fusion-rule contradiction means the headline gains are not tied to a well-defined model, but the human-study-driven framework is worth refereeing.","tokens_in":16768,"tokens_out":2897,"would_cite":false,"duration_ms":30411,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that multi-face deepfake detection improves when the detector is built around four contextual cues humans actually use: scene-motion coherence, inter-face appearance compatibility, interpersonal gaze alignment, and…","keywords":["deepfake detection","multi-face detection","human cognition","contextual cues","frame-level complete detection","gaze alignment","face-body consistency","interpretability"],"falsifier":"Run a controlled study in which hundreds of naive viewers are shown newly generated multi-face deepfakes not drawn from the benchmark datasets and asked to state why faces look fake; if the four cues are not the dominant reasons they give, or if a version of HICOM with the gaze and body modules ablated matches the full model on those videos, the central claim is falsified.","tokens_in":15828,"feed_emoji":"🎭","tokens_out":9108,"duration_ms":96213,"temperature":0.7,"pith_summary":"The paper sets out to show that detecting every fake face in a multi-face video is a different task from detecting one fake face, and that humans solve it by reading social context rather than inspecting faces in isolation. Through human studies, it identifies four cues people rely on: whether a face moves coherently with the scene, whether it matches the other faces in appearance, whether its gaze fits the group's gaze pattern, and whether its face agrees with its body in age and gender. The proposed system, HICOM, turns each cue into a module and fuses their outputs so that a frame is judged correct only if every face in it is classified correctly. On standard multi-face benchmarks the paper reports average frame-level accuracy gains of 3.3% in-dataset and 5.8% on an unseen dataset, plus human-readable explanations from a language model.","feed_headline":"Human-inspired detector lifts multi-face deepfake accuracy by 5.8%","feed_subtitle":"A four-module system uses gaze, motion, appearance, and body cues to flag every manipulated face, not just one.","key_machinery":"The central object is HICOM, a four-module detector whose modules mirror the four human cues. M1 infers scene-motion coherence from multi-scale face and background features over time; M2 compares faces within a frame using a Transformer trained with contrastive and cross-entropy losses; M3 trains a gaze-locking classifier on eye regions and flags off-camera faces only when most faces in the scene face the camera; M4 trains age and gender classifiers separately on face crops and body crops and flags mismatches. A final XOR fusion marks a face as fake if any module flags it, so a single working cue can still catch a manipulated face.","core_discovery":"The central claim is that contextual inconsistency, not individual facial artifact, is the reliable signal for catching deepfake faces in social scenes. A two-phase human study yields four dominant cues, and HICOM encodes them as four modules: scene-motion coherence, inter-face appearance comparison, interpersonal gaze alignment, and face-body age and gender consistency. The paper reports that existing methods saturate on face-level accuracy but drop on frame-level complete multi-face detection, while HICOM improves average frame-level accuracy by 3.3% across FFIW, OpenForensics, and DF-Platter, maintains an edge under unseen perturbations, and generalizes to the unseen ManualFake dataset with a 5.8% average frame-level accuracy improvement.","pith_inferences":["Editorial inference: if the four cues are genuinely stable across deepfake generations, detectors built on them should degrade more gracefully on future generation methods than detectors trained on pixel-level artifacts.","Editorial inference: the gaze module's decision rule presupposes that most faces in a scene look at the camera; in candid or surveillance footage where nobody does, this cue goes silent, so the framework would need a different social baseline for such scenes.","Editorial inference: the paper's cue-prevalence numbers could be tested against learned fusion weights, which would reveal whether the human-reported proportions (motion 34.2%, appearance 31.5%, gaze 25.0%, body 7.5%) are also the optimal ones for the model."],"forward_implications":["Frame-level complete multi-face detection becomes a metric that separates methods, since face-level accuracy alone is nearly saturated across baselines.","Human-inspired contextual cues transfer to unseen datasets and perturbations better than artifact-specific features, as shown by the 5.8% gain on the untrained ManualFake dataset.","The modular design means a newly discovered human cue can be added as an extra module without retraining the entire system.","XOR fusion makes detection robust in the sense that one reliable cue is sufficient to flag a fake face even when the other modules fail.","An attached language model can turn module scores into plain-language explanations for each face, making the system's verdicts inspectable by non-experts."],"supporting_citations":[{"why":"Supplies the OpenForensics multi-face forgery dataset used in the human study and as an in-dataset evaluation benchmark.","marker":"[39]"},{"why":"Supplies the FFIW real-world multi-face deepfake video benchmark used for human study and evaluation.","marker":"[84]"},{"why":"Supplies the DF-Platter multi-face deepfake dataset used in the human study and evaluation.","marker":"[53]"},{"why":"Supplies ManualFake, the unseen dataset on which cross-dataset generalization is tested.","marker":"[26]"},{"why":"Provides a recent multi-face detection baseline whose reported numbers serve as comparison points in the tables.","marker":"[43]"},{"why":"Provides a recent multi-face forgery detection baseline used for comparison.","marker":"[81]"},{"why":"Provides the multi-face detection method whose evaluation protocol on ManualFake the paper follows.","marker":"[50]"},{"why":"Supplies the gaze dataset used to pretrain the gaze-locking model in module M3.","marker":"[69]"},{"why":"Supplies the age-and-gender-labeled face dataset used to train the age and gender models in module M4.","marker":"[63]"}],"fun_headline_variants":["Human-inspired cues spot every fake face in groups","Multi-face deepfake detection boosts accuracy by 5.8%","Seeing like humans: new detector catches all deepfake faces","Gaze, motion, and body cues unmask deepfake groups","Context beats artifacts: human-inspired deepfake detector"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim depends on the assumption that the cues identified by a small group of human viewers on the benchmark datasets are the stable, transferable signals that drive detection, rather than quirks of those particular videos or viewers.","fun_headline_variants_meta":{"raw":{"variants":["Human-inspired cues spot every fake face in groups","Multi-face deepfake detection boosts accuracy by 5.8%","Seeing like humans: new detector catches all deepfake faces","Gaze, motion, and body cues unmask deepfake groups","Context beats artifacts: human-inspired deepfake detector"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000166,"raw_usage":{"total_tokens":1244,"prompt_tokens":927,"completion_tokens":317,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":543,"completion_tokens_details":{"reasoning_tokens":236}},"tokens_in":543,"tokens_out":317,"duration_ms":3956,"temperature":1.0,"reasoning_tokens":236,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:46:51.973444+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a controlled study in which hundreds of naive viewers are shown newly generated multi-face deepfakes not drawn from the benchmark datasets and asked to state why faces look fake; if the four cues are not the dominant reasons they give, or if a version of HICOM with the gaze and body modules ablated matches the full model on those videos, the central claim is falsified.","supporting_citations":[{"cited_title":"Openforensics: Large-scale challenging dataset for multi-face forgery detection and segmentation in- the-wild","cited_arxiv_id":null,"evidence_quote":"Supplies the OpenForensics multi-face forgery dataset used in the human study and as an in-dataset evaluation benchmark."},{"cited_title":"Face forensics in the wild","cited_arxiv_id":null,"evidence_quote":"Supplies the FFIW real-world multi-face deepfake video benchmark used for human study and evaluation."},{"cited_title":"Df-platter: Multi- face heterogeneous deepfake dataset","cited_arxiv_id":null,"evidence_quote":"Supplies the DF-Platter multi-face deepfake dataset used in the human study and evaluation."},{"cited_title":"Exploiting facial relationships and feature aggregation for multi-face forgery detection.TIFS, 19:8832– 8844, 2024","cited_arxiv_id":null,"evidence_quote":"Provides a recent multi-face detection baseline whose reported numbers serve as comparison points in the tables."},{"cited_title":"Comics: End-to-end bi-grained contrastive learn- ing for multi-face forgery detection","cited_arxiv_id":null,"evidence_quote":"Provides a recent multi-face forgery detection baseline used for comparison."},{"cited_title":"Mixture-of-Noises Enhanced Forgery-Aware Predictor for Multi-Face Manipulation Detection and Localization","cited_arxiv_id":"2408.02306","evidence_quote":"Provides the multi-face detection method whose evaluation protocol on ManualFake the paper follows."},{"cited_title":"Gaze locking: passive eye contact detection for human- object interaction","cited_arxiv_id":null,"evidence_quote":"Supplies the gaze dataset used to pretrain the gaze-locking model in module M3."},{"cited_title":"Dex: Deep expectation of apparent age from a single image","cited_arxiv_id":null,"evidence_quote":"Supplies the age-and-gender-labeled face dataset used to train the age and gender models in module M4."}],"review_version":1}