- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper presents VeilTrace, a recoverable image-hiding framework for remote face-forgery detection and artifact localization. It conceals the input face in a natural cover image and reconstructs it at an authorized endpoint for forensic analysis. The method introduces a consistency loss over fixed forensic responses to reduce changes in local responses during reconstruction. It also proposes Hybrid-DSSIM, which selects full or mask-restricted localization targets based on the geometric reliability of pristine–forged pairs. The evaluation includes cross-dataset detection, localization ablations, unauthorized-recovery attacks, and transmission distortions. The reported results indicate that detection performance is largely preserved under clean transmission and that Hybrid-DSSIM improves spatial correspondence between localization predictions and supervision targets.
Critical strengths
1.Preservation of downstream detection performance under clean transmission. In Table 2, the main FA+SBI setting on recovered images and the corresponding plaintext baseline both achieve a rounded mean frame-level AUC of 88.98%. This provides empirical support for using conventional image-domain detectors after authorized recovery and substantiates the feasibility of the proposed framework.
2.Targeted treatment of geometric noise in localization supervision. Hybrid-DSSIM selects supervision targets according to pairwise geometric reliability and provides a mask-supported fallback for unreliable pairs. The controlled comparison fixes the feature encoder and diffusion architecture, while Table 7 further shows greater sample-specific correspondence gain than the alternative supervision strategies. These results support the value of the target-construction strategy for localization training with imperfectly aligned pristine–forged pairs.
Critical weaknesses
1.I have reservations about the interpretation of the forensic-response consistency ablation. In Table 1(b), adding this loss increases mean AUC from 80.26% to 81.58%, but also improves recovery PSNR from 49.06 dB to 49.76 dB. Although the ablation demonstrates a benefit, it does not fully isolate the effect of forensic-response preservation from general improvements in reconstruction quality. Since this loss is a central technical contribution, a comparison at similar reconstruction quality, or against alternative reconstruction losses, would strengthen the evidence for additional forensic benefits.
2.The empirical evidence for privacy protection remains limited. Table 5 evaluates unauthorized recovery using reconstruction PSNR, SSIM, RGB correlation, and face-detection rates. These metrics do not directly assess whether a keyless observer can infer identity information. Appendix C.3 also explicitly notes that face detection is not an identity-leakage measure. Although the authors appropriately limit their security claims, the current results leave the extent of identity protection unclear. Given that privacy preservation is a central motivation of the paper, a direct identity-matching or identity-inference evaluation under the stated threat model would provide stronger evidence for the protection afforded by the framework.
Questions for authors
1.Does the detection gain from the forensic-response consistency loss persist at comparable reconstruction quality? Could the authors provide a controlled comparison using the same hiding architecture and detector-training setup, with similar recovery PSNR/SSIM, or a comparison against alternative reconstruction losses? This would help determine whether the gain in Table 1(b) reflects forensic-response preservation beyond general reconstruction improvements.
2.Have the authors directly evaluated whether a keyless observer can infer or match identities from protected images or unauthorized reconstructions under the stated threat model? If available, could they report the evaluation setup and results? Since reconstruction fidelity and face-detection rates are not direct measures of identity leakage, such evidence would help clarify the extent of privacy protection provided by the framework.
2 comments
Thank you for the additional experiments. The comparison in Table 6 answers my first question. At similar recovery PSNR and SSIM, the forensic-response consistency loss better preserves the predefined responses, but this comparison does not show an additional gain in detection AUC. I appreciate the authors’ clarification of what the loss was shown to achieve.
The new privacy experiments also help clarify the scope of the protection. The tested keyless reconstruction method did not recover clear faces, but when the original cover image was known, source-video matching reached an AUC of 0.980. As you note, this measures linkage to a source video and leaves open whether a person’s identity can be inferred. I would encourage you to make this distinction explicit in the paper.
The comparison using the same detector checkpoint and the held-out localization results strengthen the evidence that the framework can support forensic analysis when the protected image is transmitted unchanged. However, Table 4 of the original submission shows that detection AUC after recovery drops to around 50% under the tested JPEG compression, resizing, and cropping conditions. The new steganalysis results also show that protected images are readily distinguishable from ordinary cover images in the tested setting. Given these findings, could you clarify in which practical forensic workflows image hiding offers an advantage over encrypting the face in transit and decrypting it at the same trusted endpoint?
The additional experiments strengthen the evidence for feasibility under clean transmission, but the practical advantage of the hiding approach remains unclear to me. I am keeping my rating at 2 (weak rejection).
Important Points (relevant to Reviewer X8h7)
IP2. Security across attacker capabilities. We evaluate carrier detection, keyless face reconstruction, and feature-based source matching (Tables 1–3). Although steganalyzers can detect protected carriers, this does not directly recover the hidden face. Reconstructions under the tested U-Net attacks achieve about 12.6 dB PSNR. We further contrast source matching under protected-only and known-cover observations to quantify how attacker knowledge affects linkage.
IP3. Controlled detection and held-out localization. We evaluate paired original and recovered frames with the same frozen FA checkpoint and video-clustered confidence intervals. Their mean AUCs are 88.982% and 89.027% (Table 4). Four frozen localizers are tested on 22388 official FF++ images. Hybrid-DSSIM achieves the highest target correlation and pair-specific gain on the reported 8551 geometry-fallback cases (Table 5).
Table 1. Steganalysis (1200 test pairs)
Table 2. Keyless reconstruction (1200 tests)
Table 3. FF++ source-video matching (not person ID)
Table 4. Same-checkpoint frame AUC (%) and paired CIs
Table 5. FF++ test, 8551 geometry-fallback cases
Reviewer X8h7
W1/Q1. Does Lfor help beyond reconstruction quality? We compare the full model with a no-Lfor checkpoint selected to match recovery PSNR/SSIM on 800 validation images, using the same hiding architecture and downstream protocol, without selecting checkpoints by test AUC. At matched quality (Table 6), the full model retains higher fixed forensic-response correlation (0.9598 vs. 0.9473, Δ = +0.0125, 95% CI [0.0122, 0.0128]). Detection AUC is comparable (98.42% vs. 98.55%, Δ = −0.121 pp, CI [-0.259, 0.005]). The closely matched reconstruction quality rules out a simple PSNR-based explanation for the response-correlation difference. Thus Lfor preserves its predefined forensic responses beyond RGB quality, without establishing a separate detection-AUC gain. CIs use 5000 video-clustered draws over 578 videos.
W2/Q2. Can a keyless observer infer identity? We test direct recognition and learned matching rather than infer privacy from reconstruction PSNR. ArcFace detects faces in 11.4% of protected probes and none of the U-Net outputs. For 4000 test pairs, a metric CNN achieves AUC 0.664 and Rank-1 1.61% on protected images, increasing to 0.980 and 33.12% when xc is known. This identifies known-cover source-video linkage as the stronger tested threat. ArcFace requires detectable facial structure, whereas the metric CNN can learn associations directly from protected inputs. FF++ source-video labels establish source association, not verified person identity.
Table 6. Quality-matched forensic-loss comparison