- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper presents ForensicZoom, a multimodal large language model framework for industrial face-forgery detection. Starting from Qwen3-VL-8B-Instruct, it progressively (i) adapts the upper vision-encoder layers and adds a Forensic Query Bank, (ii) aligns visual representations with the language model through supervised fine-tuning, (iii) trains a PPO-based NEED_ZOOM policy that optionally triggers a second inference pass with an attention-guided enlarged crop, and (iv) refines natural-language attributions through cascade DPO/GSPO optimization. On a proprietary identity-verification dataset of approximately 2.05 million samples, the paper reports 97.3% TPR at 0.1% FPR on AIGC forgeries, compared with 91.0% for the strongest specialized baseline reported in the paper. It also discusses inference latency and an asynchronous deployment pipeline.
Critical strengths
<1>The paper targets identity-verification security rather than an unconstrained image classification task. The sizable industrial dataset, user-disjoint split, and emphasis on TPR at 0.1% FPR make the evaluation setting practically meaningful. The reported 97.3% TPR is promising, although the private-data setting limits independent verification. <2>The four-stage design connects task-specific perception, language alignment, adaptive inspection, and attribution optimization. Stage-wise ablations and the reported single-/two-pass latency and manual-review efficiency provide a useful account of the accuracy–inspection-cost trade-off, even though the causal benefits of the zoom policy remain to be isolated.
Critical weaknesses
<1>Stage 3 raises TPR from 93.7% to 96.9%, but simultaneously changes the learned policy, available visual evidence, and number of inference passes. The paper does not compare against a second pass using only the original image, random cropping, always zooming, MLLM-predicted regions, or confidence/entropy/logit-margin triggering. Consequently, it is unclear whether the gains arise from meaningful region selection and a learned when-to-zoom policy, rather than repeated inference or a simpler strategy. These omissions directly affect the paper’s central claim.
<2>The method uses the final-layer ViT attention map to choose a crop and aligns a text-derived spatial distribution to that map. Yet a concentrated attention map is not necessarily a faithful localization of manipulation evidence, and text–attention consistency alone does not guarantee factual grounding. The paper does not adequately establish its distinction from related active-visual-inspection and tool-use work, including V*, Active-o3, and ForgeryVCR. Furthermore, detection is evaluated only on proprietary data, without a public benchmark or convincing unseen-attack/temporal-shift evaluation; uncertainty estimates and threshold-calibration details at the stringent operating FPR are also missing. Together, these issues make both the methodological contribution and its generality difficult to assess.
Questions for authors
Q1. Is zooming itself necessary? With the same Stage 2 model and a matched inference budget, what are the results for (a) a second inference pass using the original image only, (b) a random 48×48 crop, (c) always applying attention-guided zoom, and (d) the proposed adaptive zoom policy?
Q2. Are the when/where policies preferable to simple alternatives? At matched zoom rates, how does PPO-based triggering compare with confidence-, entropy-, or logit-margin thresholds? How does attention-guided crop selection compare with directly having the MLLM predict the suspicious bounding box?
Q3. What validates the attention-based reference? What supports the assumption that the last-layer attention map corresponds to actual manipulation regions? In particular, what evidence shows that CoT Attention Alignment improves the factual grounding of the textual region description rather than simply making two model-generated signals agree?
Q4. How robust are the low-FPR results? Can the authors provide public-benchmark or held-out-generator/temporal-shift results, uncertainty estimates, and the precise protocol for selecting and fixing the 0.1%-FPR operating threshold?
Q5. How does the method differ from the closest prior approaches? Please identify the specific methodological improvements over V*, Active-o3, and especially ForgeryVCR, and position the work relative to relevant agentic-forensics and active-visual-search literature (e.g., AIFo, ForenAgent, DeepEyes, and Chain-of-Focus).