acceptodds
Under review as a conference paper at ICLR 2027

MEGRA: Mixture-of-Experts and Graph Retrieval-Augmented Generation for Explainable Face Forgery Detection

Abstract

Conventional deepfake detection methods typically output only authenticity labels, forgery scores, or visual heatmaps, making it difficult to provide natural-language explanations that are readily understandable to humans. Recent methods based on multimodal large language models (MLLMs) can generate natural-language explanations, offering a new direction for explainable deepfake detection. However, they still face two challenges. First, their modeling of fine-grained forgery cues remains insufficient, resulting in poor generalization to forgery mechanisms unseen during training. Second, MLLM-generated explanations are prone to hallucination. To mitigate these issues, we propose MEGRA (Mixture-of-Experts and Graph Retrieval-Augmented Generation), a face forgery detection method that combines mechanism-aware Mixture-of-Experts (MoE) detection with Graph Retrieval-Augmented Generation (GraphRAG) explanation. MEGRA comprises two modules: a forgery detection module and an explanation module. In the forgery detection module, we design a mechanism-oriented Forgery Detection MoE (FD-MoE) that adaptively combines experts with complementary forensic capabilities to better discriminate among different forgery mechanisms and adapt to distribution shifts. In the explanation module, we construct face-Forensic Knowledge Graphs (F-FKGs) and introduce GraphRAG to retrieve relevant forensic knowledge conditioned on the detection result and regional forensic evidence, guiding the MLLM to generate evidence-grounded natural-language explanations. Experiments on HydraFake show that MEGRA achieves an average accuracy of 91.09%, reaching state-of-the-art (SOTA) performance. In addition, MEGRA’s explanations receive the highest scores among reported results under all three judge models: GPT-5.5, Kimi-K3, and Qwen3.7-Plus.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
deepfake detection, explainable AI, mixture of experts, GraphRAG, multimodal large language models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_CD8uOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes MEGRA, an explainable face forgery detection framework combining a mechanism-aware Mixture-of-Experts (MoE) detector with Graph Retrieval-Augmented Generation (GraphRAG). The detection module integrates multiple forensic experts, while the explanation module uses structured forensic knowledge to generate natural-language explanations. The method achieves competitive detection performance on HydraFake and improves explanation scores evaluated by multiple MLLM judges.

Critical strengths
  1. The paper addresses the practically relevant problem of combining deepfake detection with human-readable forensic descriptions.

  2. The experiments cover multiple out-of-distribution scenarios, and the proposed method demonstrates improvements on several challenging cross-domain subsets.

  3. The authors provide component-level ablations for both detection and explanation modules.

Critical weaknesses
  1. Limited methodological novelty.

The main components, including MoE-based detector fusion, retrieval-augmented forensic reasoning, and MLLM-based explanation generation, have already been extensively explored in recent literature. The proposed framework primarily combines these established techniques without introducing a sufficiently distinctive learning mechanism or conceptual insight. Given the substantial number of related works published in 2024–2026, the incremental contribution appears insufficient for ICLR.

  1. The claimed explainability is not adequately justified.

A fundamental concern is what the proposed framework actually explains. Generating plausible forensic descriptions is not equivalent to explaining why a detector makes a particular prediction. In MEGRA, the regional evidence branch does not participate in the detection decision, and the explanation module operates on a fixed prediction. Consequently, there is no clear evidence that the generated explanations faithfully reflect the detector's actual decision-making process.

The authors should distinguish between evidence-grounded description and faithful model explanation. High scores from MLLM judges do not establish explanation faithfulness.

  1. Insufficient evidence for the necessity of GraphRAG.

The ablation experiments demonstrate that incorporating forensic knowledge improves explanation scores. However, they do not establish whether graph-based retrieval is necessary. Comparisons against conventional RAG and structured knowledge prompting under controlled conditions are missing. Therefore, the specific methodological contribution of GraphRAG remains unclear.

  1. Modest overall improvement relative to system complexity.

MEGRA improves the average detection accuracy from 90.70% to 91.09% over VERITAS. Although the cross-domain improvements are more noticeable, the overall gain is relatively small considering the integration of multiple pretrained detectors, mechanism-specific experts, calibration components, and additional explanation machinery. It remains unclear whether the improvements result from a novel learning strategy or primarily from combining complementary existing models.

Questions for authors

See the weaknesses above.

Reviewer_pcB4OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, the authors propose a mixture-of-experts architecture for deepfake detection and localization. For detection, the authors combine the prediction scores of six experts and introduce a score calibration scheme to align their outputs. For localization, they employ the Xception to generate an evidence map, which is combined with rules and a knowledge graph to reduce hallucinations in LLM-generated explanations. Experiments demonstrate the effectiveness of the proposed method.

Critical strengths
  1. The MoE architecture is a meaningful direction for generalizable deepfake detection.

  2. The proposed method achieves state-of-the-art performance.

Critical weaknesses
  1. The criteria for selecting experts are not clear. First, the manuscript does not report error overlap between experts, which experts correct others’ mistakes, or the benefit of adding each expert under matched conditions. Moreover, removing SBI and Xception reduces accuracy only from 91.09% to 90.54%, suggesting a marginal contribution that does not strongly support the need for all six experts. Furthermore, UCF and RECCE are both Xception-based detectors, while UnivFD and Effort share the CLIP ViT-L/14 backbone. These shared architectures raise concerns about the effective diversity of the expert pool.

  2. Why is Xception, whose standalone detection performance is limited, selected as the source of regional evidence? Xception's detection AUC falls below 70% in many practical settings. Is it reliable to use its responses as the only source of localization cues? In particular, are these regional responses reliable for forgery mechanisms unseen during training? The manuscript should justify this choice and directly evaluate the quality of the resulting regional evidence.

  3. How is the knowledge graph constructed? The authors describe its structure but do not clearly explain where the knowledge comes from. Specifically, i) which publications, databases, or expert records provide the supporting text? ii) Are the 88 rules manually written, model-generated, or automatically extracted from the literature? iii) Who reviews these rules, and what expertise do they have? iv) How are conflicts, duplicates, applicability conditions, and exceptions handled? v) What evidence establishes that a particular cue supports a given forgery category? Furthermore, have the correctness and coverage of the graph been independently evaluated? vi) If the graph needs human verification, is there a proper IRB?

  4. The explanation evaluation mixes detection correctness with explanation quality. The scoring rubric assigns a score of 1 to an incorrect authenticity judgment, so a better detector can receive higher explanation scores without producing more faithful explanations. Moreover, overall scores for detail, clarity, and comprehensiveness do not directly measure hallucinations. The paper does not report the correctness of individual visual claims or the frequency of unsupported cues and incorrect forgery-type attributions.

Questions for authors

Please see weaknesses. I will re-consider the rating if the concerns are resolved.

Reviewer_UoJ9OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper targets explainable face forgery detection, arguing that existing MLLM-based detectors (i) model fine-grained forgery cues insufficiently, hurting generalization to unseen mechanisms, and (ii) hallucinate in their natural-language explanations. It proposes MEGRA, which couples a mechanism-aware Mixture-of-Experts detector (frozen forensic experts + four mechanism-specific MLP experts + class-conditional calibration + a mechanism-aware mixture network + a synthesis residual) with a GraphRAG explanation module (a shared face-forensic knowledge graph, path retrieval conditioned on decision and regional evidence, then QLoRA-tuned Qwen2.5-VL-7B). Experiments use the HydraFake ID/CM/CF/CD protocol, report 91.09% average accuracy versus 90.7% for VERITAS, and report higher LLM-judged explanation scores than five baselines under three judges.

Critical strengths

S1. The cross-domain result is real and non-trivial. On the CD split, which is the closest to open-world deployment, average accuracy improves from 82.15% to 85.74% (+3.6 points), with large per-subset gains on DeepFaceLab (80.1 vs 58.6), IC-Light (84.4 vs 75.7) and Dreamina (96.1 vs 92.3). The per-subset spread also drops (SD 6.88 vs 10.29 over the 18 OOD subsets). This is the single most defensible contribution in the paper.

S2. The evaluation protocol is adopted, not invented. The authors reuse HydraFake's official stratified split, take all non-MEGRA detection numbers directly from VERITAS to avoid reimplementation variance, and reuse VERITAS's public judge prompts and scoring rubric. This is better practice than most submissions in this area.

S3. Engineering completeness and reporting honesty. 53,155 test images, 19 evaluation subsets, calibration fitted only on validation, per-component lesion definitions (Appendix B.2), paired bootstrap and McNemar tests with Holm correction (Appendix H), reproducibility artifacts (Appendix M), and explicit reporting of two failure cases with correct decisions but incorrect forgery-type attribution (Appendix N). The paper does not hide its negative results, which makes the review tractable.

Critical weaknesses

W1. The "fine-grained" claim has no fine-grained component and no fine-grained metric. The paper motivates itself on "insufficient modeling of fine-grained forgery cues" (Sec. 1) and attributes poor generalization to that deficiency. No proposed component operates at that granularity. The four mechanism experts take the global SigLIP embedding phi(x) as input (Eq. 4), and the mixture network also estimates its coefficients from global features. The only region-aware branch is a frozen Xception sidecar over eight MediaPipe-aligned regions, and the paper states explicitly that it "does not contribute to the computation of S(x)" (Sec. 3.2.1), so it cannot affect detection. No localization metric (region IoU, pixel-level AUC, point-wise accuracy) is reported anywhere in the main paper or appendix. The diagnostic claim and the proposed mechanism therefore do not meet: whatever improves cross-domain accuracy, it is not fine-grained cue modeling as the paper defines the problem.

W2. Hallucination is declared the central problem and then never measured. "Hallucination" appears four times in the paper, entirely in the introduction and related work. No hallucination rate, faithfulness rate, unsupported-claim rate, or human evaluation is reported, and Appendix L states verbatim: "No human evaluation is reported." What actually constrains the output is an output contract, not the proposed knowledge mechanism: the decision field is copied verbatim from upstream (Eq. 2), region and cue names are drawn from a closed 9-region x 12-cue vocabulary (Appendix K.7), the forgery type is one of four fixed closing sentences (Table 22), and a repair prompt is triggered on contract failure. This is a legitimate design, but it means the anti-hallucination credit belongs to constrained decoding. The paper's thesis is that retrieval-grounded knowledge makes explanations reliable, and that thesis is left unmeasured. The one non-LLM faithfulness metric that is freely available, forgery-type accuracy over the closed four-way label, is not reported, even though the labels are derivable from the source-method mapping in Table 5.

W3. GraphRAG is not retrieval in any falsifiable sense, and its contribution is not identifiable. The knowledge base contains 260 nodes in total (Table 9: 21 regions, 27 cues, 88 rules, 11 categories, 3 authenticity patterns, 4 caution policies), all authored by the authors, with no external corpus and no cited provenance for the 88 rules. Retrieval is a deterministic scoring function with hand-set weights, lambda = (2.0, 1.5, 1.5, 1.0, 0.8, 0.8, 0.5) (Appendix D.4), with no learned retriever and no embedding index. The query is a function of (y-hat, eight region indices, arg-max mechanism, uncertainty), so the reachable retrieval space is finite and small. Functionally this is weighted prompt-template assembly. The naming carries a claim the mechanism does not support, and the empirical support is weaker than the name suggests. Table 4 uses one fixed SFT checkpoint and ablates only inference-time inputs, so it never tests whether training with the graph matters. Its pointwise scores are non-monotonic: full model 4.18/4.75, w/o F-FKGs 4.08/4.48, w/o F-FKGs and regions 4.12/4.57. Removing strictly more information improves the score, and the offered explanation is post hoc. The decisive missing control is a vocabulary-only condition: the graph injects exactly the 21-region and 27-cue terminology, so removing the graph also removes the vocabulary. Without that control, "structured knowledge helps" and "more forensic-sounding tokens help" are not separable, and the ELO ordering (1126.7 > 972.8 > 900.4) cannot be interpreted.

W4. The explanation evaluation cannot separate the method from the instrument. Four confounds, all verifiable in the appendix. (a) Teacher is judge: the SFT targets are generated by Kimi (Appendix K, Category D), and Kimi-K3 is simultaneously the pointwise and pairwise judge, so the student is optimized toward Kimi's style and then graded by Kimi. (b) Asymmetric same-family correction: Appendix L removes the Qwen2.5-VL-7B base from the Qwen3.7-Plus column "due to suspected same-family preference" but retains MEGRA in that column, although MEGRA is exactly Qwen2.5-VL-7B plus QLoRA; the concern applies with at least equal force to the fine-tuned model. (c) Effect size is not larger than instrument noise: the judge-to-judge spread for MEGRA (4.18 under Kimi vs 4.75 under Qwen, 0.57) is essentially the same as the claimed method gap (4.18 vs 3.59 under Kimi, 0.59). (d) The reported ELO is a deterministic monotone transform of the pairwise win rate (Appendix L: Delta = 400 log10(P/(1-P))), fitted as a star graph with MEGRA as the only hub and no baseline-baseline pairings, so 1723.3 corresponds to a 92.5% win rate against VERITAS. A 92.5% win rate and a 0.59 point difference on the same 1000 images are not mutually consistent unless the pairwise judge is responding to surface form rather than reasoning quality. The rubric compounds this: it rewards "creativity, comprehensiveness, and granularity" (Appendix K.1), and in a paper whose thesis is anti-hallucination, rewarding creativity and comprehensiveness is misaligned with the stated objective, since neither instruction penalizes unsupported claims.

W5. The paper's own ablation contradicts its central generalization claim. Removing the mechanism experts (Table 3) costs 13.3 points on ID, 10.5 on CM, 5.2 on CF, but only 1.25 on CD. The component named after mechanism-aware generalization decays to almost nothing precisely on the split that measures generalization. Conversely, the frozen experts, which alone reach 52.14-62.17% average accuracy (Table 17; six-model stack 62.93%), are the ones carrying the CD gain (removing them costs 5.8 points there). The open-world improvement is therefore delivered largely by a six-model ensemble of off-the-shelf detectors that are individually close to chance on this benchmark, whose marginal contribution is 3.58 points, while the proposed mechanism machinery contributes mostly in-domain capacity. This should be reported as the finding, not as support for the current framing.

W6. The state-of-the-art claim has no paired significance test against the only competitive baseline. The margin over VERITAS is 0.39 points on the reported average and 0.52 on OOD, while the standard deviation across the 18 OOD subsets is 6.88. Appendix H explicitly restricts its tests to lesions: "These tests concern the internal detection-module lesions only." No test of MEGRA against VERITAS is reported. The quoted confidence interval [90.77, 91.42] is computed over images rather than over subsets, so it is far narrower than a claim resting on cross-distribution generalization requires. The comparison is also asymmetric across columns: MEGRA loses on ID (95.5 vs 97.3) and by 5.2 points on FFIW (73.3 vs 78.5), and the reported average mixes denominators between rows with and without an ID entry (footnoted, but load-bearing for the SOTA sentence).

W7. The system is an ensemble compared against single models, with no cost accounting. MEGRA runs six frozen detectors, four MLP experts, a SigLIP-LoRA channel, two calibrated residual channels, an Xception region sidecar, and a 7B QLoRA-tuned MLLM with graph retrieval. Baselines in Table 1 are single models. No inference cost, parameter count, or latency comparison appears in the main paper to justify a +0.5 point OOD gain at this complexity. The 2-expert variant collapsing to 75.92 (Table 18), while the uniform average of the same experts rises, further suggests the stacking head is highly sensitive to pool composition, which weakens the mechanism-aware interpretation of the gain.

W8. No limitations section. Appendix N is titled "Additional Qualitative Examples and Limitations" but contains no limitations.

Questions for authors

Q1 (W1). Which specific component models fine-grained forgery cues, and on what metric is its fine-grained behavior verified? Region-level localization accuracy or pixel-level AUC on a public localization benchmark would settle this. Relatedly, since the Xception region branch does not enter S(x), what is the evidence that fine-grained cue modeling is what drives the CD gain?

Q2 (W2). What is the forgery-type accuracy over the closed four-way output label (Table 22), using the source-method to FS/FR/EFG mapping in Table 5? This is the only faithfulness metric here that does not depend on an LLM judge, and its absence is conspicuous given that hallucination is the paper's stated motivation. Please also report the rate of outputs that fail contract validation and require the repair prompt.

Q3 (W3). Please report a vocabulary-only control: the same nine-region and twelve-cue output contract and the same retrieved-path slot length, but with generic forensic text replacing F-FKG paths, with the SFT model retrained under that condition. Does the pointwise advantage and the ELO ordering survive? This is the minimum required to attribute the explanation gain to graph structure rather than to forensic terminology.

Q4 (W4). Please rescore all baselines with the ground-truth decision supplied as input, so that the explanation comparison is no longer confounded with detection accuracy; as written, the rubric assigns score 1 when "the answer is incorrect," which structurally advantages the method with the strongest detector on a scale presented as measuring explanation quality. Please also rerun Table 2 with a judge from a different family than both the teacher and the student.

Q5 (W6). Please report a paired test of MEGRA against VERITAS on the same images with the subset, not the image, as the unit of resampling. Given a 0.52 point OOD margin against a 6.88 point cross-subset standard deviation, what is the basis for describing the result as state of the art?

Q6 (W7). What is the marginal cost, in parameters and latency, of the explanation module and of the six-experts-plus-four-MLPs detection module relative to a single SigLIP-LoRA classifier trained on the same data? Since removing the frozen experts still leaves 87.51, a single-model baseline under the identical training recipe would clarify how much of the reported accuracy is attributable to supervision on HydraFake rather than to the proposed architecture.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.