acceptodds
Under review as a conference paper at ICLR 2027

Privacy-Preserving Deepfake Detection and Fine-Grained Artifact Localization with Forensic-Aware Image Hiding

Abstract

Remote face-forgery analysis typically requires transmitting identifiable facial content to remote services, creating privacy risks during transmission and storage. Encryption can conceal image content, but an image-like ciphertext may reveal the presence of protected communication and facilitate traffic-level discovery or tracking. Moreover, a conventional decrypt-then-infer pipeline exposes plaintext at the endpoint that performs inference. We propose VeilTrace, a recoverable image-hiding framework that represents a face as a cover-like protected image and reconstructs it only at an authorized forensic endpoint for downstream detection and fine-grained localization. VeilTrace addresses two key challenges. First, high RGB reconstruction fidelity does not guarantee preservation of forensic signals. We therefore introduce a forensic-response consistency objective based on fixed forensic proxy responses. Second, residual misalignment between real/fake pairs can contaminate pixel-level DSSIM supervision. We therefore apply full DSSIM only to geometrically reliable pairs and use a mask-supported fallback for unreliable pairs when a usable manipulation mask is available, yielding our Hybrid-DSSIM strategy. Extensive experiments show that VeilTrace reduces direct exposure of facial plaintext under the evaluated threat model while retaining high forgery-detection accuracy, with a rounded mean frame-level AUC of 88.98% in the FA+SBI setting and 49.76 dB recovery PSNR. Hybrid-DSSIM yields the strongest artifact-map agreement in the controlled ablation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Deepfake Detection, Privacy-Preserving Computer Vision, Image Hiding, Face Forgery Localization, Digital Forensics

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_X8h7OpenReview ↗
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

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)

MethodAccuracy (%)ROC-AUCTPR@1% FPR
SRM + FLD99.581.0001.000
CRM + FLD100.001.0001.000
SRNet100.001.0001.000

Table 2. Keyless reconstruction (1200 tests)

InputPSNR (dB)SSIM
Protected xtx_t12.610.368
Known cover [xt,xc,xt−xc][x_t,x_c,x_t-x_c]12.580.364

Table 3. FF++ source-video matching (not person ID)

Probe / methodFace cov.AUCTPR@1%R1 (%)
ArcFace / protected11.4%0.4660.0000.00
ArcFace / U-Net output0%–––
Metric CNN / protectedn/a0.6640.0271.61
Metric CNN / known covern/a0.9800.72133.12

Table 4. Same-checkpoint frame AUC (%) and paired CIs

DatasetRawRec.Δ [95% CI], pp
CDF-v187.9087.65−0.25 [−1.27, 0.72]
CDF-v288.4888.34−0.14 [−0.56, 0.27]
DFD93.4893.31−0.17 [−0.31, −0.03]
DFDC85.7385.56−0.16 [−0.31, −0.01]
DFDCP90.9991.10+0.11 [−0.22, 0.46]
FaceShifter80.9780.90−0.07 [−0.57, 0.48]
UADFV95.3396.33+0.99 [0.29, 1.86]
Unweighted mean88.98289.027+0.045

Table 5. FF++ test, 8551 geometry-fallback cases

TargetCore PSNRCorr.Pair gain
Naive DSSIM13.990.2740.027
Valid only15.380.4870.041
Mask only8.980.6650.042
Hybrid-DSSIM17.040.6820.112

Reviewer X8h7

W1/Q1. Does LforL_{for} help beyond reconstruction quality? We compare the full model with a no-LforL_{for} 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 LforL_{for} 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 xcx_c 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

MetricW/oFullMatched
PSNR (dB)48.4349.0749.10
SSIM0.996670.997120.99711
Forensic corr.0.94170.95980.9473
AUC (%)98.5298.4298.55

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).

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

VeilTrace hides a face in a natural-looking cover image, reconstructs it at an authorized endpoint, and then runs standard deepfake detection and diffusion-based localization. It uses a nine-channel forensic-response loss to preserve low-level traces during recovery. Hybrid-DSSIM builds localization targets from aligned DSSIM maps when alignment is reliable and from manipulation masks otherwise.

Critical strengths
  1. A useful end-to-end problem. The paper combines recoverable hiding, deepfake detection, and spatial localization in one pipeline. Preserving forensic evidence through hiding and recovery is a meaningful engineering goal.
  2. Clear failure cases. The authors report that JPEG 70, resizing, and cropping reduce reconstruction to about 12–13 dB and detection AUC to near chance. These results make the method's narrow operating range easier to judge.
Critical weaknesses
  1. The privacy evidence is too weak for the main claim. VeilTrace reconstructs the full face at the authorized endpoint, so privacy depends on that endpoint being trusted. The paper does not explain what protection this adds over encrypted transport. PSNR/SSIM, RGB correlation, and face-detection rate do not measure steganographic detectability or identity leakage, and the evaluation lacks adaptive white-box attacks, face verification, and cross-frame linkability tests. The current results show recoverable cover-like transport under limited conditions, not general privacy preservation.
  2. Detection and localization are not evaluated under a common protocol. It is unclear whether Raw and recovered images use the same detector checkpoint.
Questions for authors
  1. Do the Raw and recovered rows in Table 2 use the same detector weights? Please evaluate both with one raw-trained, frozen detector and report confidence intervals clustered by video or identity.
  2. Are Tables 3 and 7 evaluated on a test set untouched by early stopping?
1 comment

Important Points (relevant to Reviewer vknU)

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)

MethodAccuracy (%)ROC-AUCTPR@1% FPR
SRM + FLD99.581.0001.000
CRM + FLD100.001.0001.000
SRNet100.001.0001.000

Table 2. Keyless reconstruction (1200 tests)

InputPSNR (dB)SSIM
Protected xtx_t12.610.368
Known cover [xt,xc,xt−xc][x_t,x_c,x_t-x_c]12.580.364

Table 3. FF++ source-video matching (not person ID)

Probe / methodFace cov.AUCTPR@1%R1 (%)
ArcFace / protected11.4%0.4660.0000.00
ArcFace / U-Net output0%–––
Metric CNN / protectedn/a0.6640.0271.61
Metric CNN / known covern/a0.9800.72133.12

Table 4. Same-checkpoint frame AUC (%) and paired CIs

DatasetRawRec.Δ [95% CI], pp
CDF-v187.9087.65−0.25 [−1.27, 0.72]
CDF-v288.4888.34−0.14 [−0.56, 0.27]
DFD93.4893.31−0.17 [−0.31, −0.03]
DFDC85.7385.56−0.16 [−0.31, −0.01]
DFDCP90.9991.10+0.11 [−0.22, 0.46]
FaceShifter80.9780.90−0.07 [−0.57, 0.48]
UADFV95.3396.33+0.99 [0.29, 1.86]
Unweighted mean88.98289.027+0.045

Table 5. FF++ test, 8551 geometry-fallback cases

TargetCore PSNRCorr.Pair gain
Naive DSSIM13.990.2740.027
Valid only15.380.4870.041
Mask only8.980.6650.042
Hybrid-DSSIM17.040.6820.112

Reviewer vknU

W1. Broader privacy tests. Beyond PSNR and face-detection rate, we evaluate three distinct attacker goals: detecting a protected carrier (Table 1), reconstructing its hidden face with or without xcx_c (Table 2), and matching source videos using learned features (Table 3). These tests distinguish visible carrier detection, image reconstruction and learned source linkage, which may succeed even when reconstruction fails.

W2/Q1. Identical detector weights. Yes. We use a single frozen FA checkpoint trained on raw frames for both original and authorized-recovered frames, with paired samples and identical preprocessing, and without retraining on recovered data. The mean frame AUC is 88.982% versus 89.027% (+0.045 pp). Table 4 reports paired differences with video-clustered 95% CIs for each dataset. The comparison therefore isolates the effect of authorized recovery from any change in detector weights.

Q2. Held-out localization. Original Tables 3(b) and 7 used an internal validation split, while Table 3(a) includes cross-protocol literature results. We therefore test four frozen localizers on the official FF++ test split (22388 images), unused for localizer checkpoint selection. Table 5 reports the 8551 geometry-fallback cases, where inaccurate pair alignment makes ordinary DSSIM supervision unreliable. Hybrid-DSSIM yields the highest core PSNR, correlation and pair-specific gain. This gain subtracts correlation with a shuffled same-type target from correlation with the paired target, testing whether the predicted map follows the actual pair rather than a generic manipulation pattern. The reported metrics measure continuous-target agreement, not binary-mask IoU.

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

This paper proposes VeilTrace, a recoverable image-hiding framework for privacy-preserving deepfake detection and fine-grained forgery localization. The proposed method employs a key-based image-hiding mechanism to embed private facial images into natural cover images. The original facial images are subsequently recovered at an authorized forensic endpoint for downstream deepfake detection and forgery localization.

Critical strengths

The paper highlights that high RGB reconstruction fidelity does not necessarily guarantee the preservation of high-frequency textures, edges, and local forgery artifacts that are critical for deepfake detection. The proposed forensic-response consistency objective further improves forensic-response correlation and downstream detection AUC while maintaining high image reconstruction quality. This suggests that introducing feature-preservation constraints tailored to forensic tasks is well motivated and offers practical value.

Critical weaknesses
  1. The paper identifies the need to recover plaintext at the inference endpoint in conventional decrypt-then-infer pipelines as one of the limitations of encryption-based forensics. However, VeilTrace itself also requires the original facial images to be recovered at an authorized forensic endpoint before deepfake detection and forgery localization can be performed. If the forensic endpoint is trusted, a conventional encryption scheme could likewise perform decryption exclusively at that endpoint. Conversely, if the endpoint is untrusted, VeilTrace cannot prevent the exposure of recovered facial plaintext either. Therefore, in terms of endpoint plaintext exposure, there appears to be no fundamental distinction between the two approaches that would justify the emphasis placed on this aspect in the paper.

  2. The paper repeatedly emphasizes that VeilTrace conceals private facial images within natural cover images, thereby addressing the limitation of conventional ciphertext potentially revealing the presence of protected communication. However, the current experiments primarily use PSNR and SSIM to measure the visual similarity between protected images and their original cover images, without systematically evaluating whether an attacker can detect the presence of hidden information. In fact, visual similarity and statistical indistinguishability are two distinct properties. Even when a protected image achieves high PSNR relative to its cover image, steganalysis detectors based on image residuals, higher-order statistical features, or deep learning may still be able to distinguish between them.

Questions for authors
  1. Both VeilTrace and conventional decrypt-then-infer pipelines require the original facial images to be recovered at an authorized forensic endpoint before detection can be performed. Therefore, there appears to be no fundamental difference between the two approaches regarding plaintext exposure at the trusted endpoint. Could the authors clarify what additional security or privacy properties VeilTrace provides compared with a conventional encryption scheme in which decryption and inference are performed at the same trusted forensic endpoint? If the primary advantage is communication covertness rather than endpoint privacy, please clarify the corresponding threat model, protection objectives, and supporting validation evidence.

  2. The paper emphasizes the visual similarity between protected images and natural cover images but lacks direct steganalysis experiments. Could the authors evaluate whether an attacker can distinguish ordinary natural cover images from VeilTrace-generated protected images and report the corresponding ROC-AUC, detection accuracy, and detection performance at a fixed low false-positive rate?

  3. Are the wrong-key/wrong-nonce tests and the learned reconstruction attacks conducted with limited training data in Table 5 sufficient to represent the capabilities of the defined unauthorized adversary? Could the authors further justify their choices of attacker training data and model capacity, and consider evaluating stronger learning-based recovery or identity-representation extraction attacks? In particular, reporting identity verification or retrieval success rates on protected images under unauthorized access would help clearly distinguish between reconstruction resistance and identity privacy.

  4. Table 5 reports the results of wrong-key, wrong-nonce, and learned unauthorized recovery experiments. The results show that incorrect keys or nonces cause the authorized decoder to fail, while the learned attacks also fail to achieve high-quality pixel-level reconstruction under the evaluated settings. However, decoding failure under incorrect credentials only demonstrates that the existing recovery function is sensitive to the supplied credentials; it does not imply that all recovery strategies operating without the correct key are incapable of extracting useful information. Although the authors additionally evaluate a learned U-Net attack, this experiment relies on a relatively limited amount of attacker training data and a single reconstruction model, which is insufficient to characterize the capabilities of stronger adaptive attackers. Furthermore, the current security evaluation primarily focuses on reconstruction PSNR, SSIM, RGB correlation, and face-detection rate. Poor pixel-level reconstruction quality does not necessarily imply that facial identity information cannot be inferred or linked.

3 comments

Important Points (relevant to Reviewer oEuk)

IP1. Why VeilTrace differs from encryption. Unlike ciphertext transport, VeilTrace transmits a cover-like RGB image without directly displaying the face, then enables key-authorized recovery compatible with existing detectors and localizers. Its advantage lies in a distinct privacy–utility trade-off for untrusted transmission and storage, rather than stronger cryptographic secrecy or endpoint confidentiality. Unlike protected-domain inference, it requires no dedicated stego-domain detector. Forensic-response consistency helps retain local artifact responses, while Hybrid-DSSIM addresses supervision corrupted by geometric misalignment (Tables 4 and 5).

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.

Table 1. Steganalysis (1200 test pairs)

MethodAccuracy (%)ROC-AUCTPR@1% FPR
SRM + FLD99.581.0001.000
CRM + FLD100.001.0001.000
SRNet100.001.0001.000

Table 2. Keyless reconstruction (1200 tests)

InputPSNR (dB)SSIM
Protected xtx_t12.610.368
Known cover [xt,xc,xt−xc][x_t,x_c,x_t-x_c]12.580.364

Table 3. FF++ source-video matching (not person ID)

Probe / methodFace cov.AUCTPR@1%R1 (%)
ArcFace / protected11.4%0.4660.0000.00
ArcFace / U-Net output0%–––
Metric CNN / protectedn/a0.6640.0271.61
Metric CNN / known covern/a0.9800.72133.12

Table 4. Same-checkpoint frame AUC (%) and paired CIs

DatasetRawRec.Δ [95% CI], pp
CDF-v187.9087.65−0.25 [−1.27, 0.72]
CDF-v288.4888.34−0.14 [−0.56, 0.27]
DFD93.4893.31−0.17 [−0.31, −0.03]
DFDC85.7385.56−0.16 [−0.31, −0.01]
DFDCP90.9991.10+0.11 [−0.22, 0.46]
FaceShifter80.9780.90−0.07 [−0.57, 0.48]
UADFV95.3396.33+0.99 [0.29, 1.86]
Unweighted mean88.98289.027+0.045

Table 5. FF++ test, 8551 geometry-fallback cases

TargetCore PSNRCorr.Pair gain
Naive DSSIM13.990.2740.027
Valid only15.380.4870.041
Mask only8.980.6650.042
Hybrid-DSSIM17.040.6820.112

Reviewer oEuk

W1/Q1. Added value over encryption. Both encryption and VeilTrace require a trusted inference endpoint. Unlike ciphertext transport, VeilTrace sends a cover-like RGB representation that does not directly display the input face and supports key-authorized recovery for existing detectors and localizers without redesigning models for a protected domain. Its contribution is this recoverable image-domain workflow with forensic-aware preservation, rather than stronger cryptographic or endpoint secrecy.

W2/Q2. Steganalysis. We train SRM–FLD, CRM–FLD and SRNet on 4000/800/1200 paired cover-stego splits separated by secret source-video ID. All three identify protected carriers with near-perfect AUC (Table 1), with TPR at 1% FPR also equal to 1. These results quantify detectability under the trained-attack protocol; detecting a carrier and recovering a face remain distinct objectives (Table 2). Under the reported test protocol, FLD uses threshold 0 and SRNet 0.5.

Q3. Stronger reconstruction attack. We expand the original attack with two independently trained U-Nets (64 base channels, 10000 updates, L1+0.2MSEL_1 + 0.2MSE) and select checkpoints by validation L1L_1. The 4000/800/1200 splits are source-video-disjoint. The attackers receive either xtx_t or [xt,xc,xt−xc][x_t,x_c,x_t-x_c]. Both yield low test recovery quality (12.61 and 12.58 dB, Table 2), even when the cover and image difference are supplied. These results characterize the tested learned attacks, rather than unrestricted white-box adversaries.

Q4. Identity-related inference. We complement reconstruction tests with direct feature-based matching (Table 3). ArcFace detects faces in only 11.4% of protected probes (59 detected positive pairs) and none of the U-Net outputs, making the latter verification AUC undefined. A metric CNN trained for source-video matching yields AUC 0.664 and Rank-1 1.61% from protected images alone. With the original cover, these rise to 0.980 and 33.12%. Thus attacker knowledge materially affects linkage, while conventional face recognition remains difficult in the tested setting. FF++ source-video labels support source association, not verified person identification. Rank-1 chance is 0.275%. TPR@1% denotes 1% FPR.

Thank you for the detailed rebuttal and the additional experiments. My main concern remains unresolved: what privacy or security benefit does VeilTrace provide over a conventional pipeline that decrypts the image before performing forensic analysis? Under the evaluated conditions, the new steganalysis results show that protected images can be detected with near-perfect performance, weakening the claim that embedding them in natural images conceals the existence of protected communication. VeilTrace also requires reconstructing the original face at a trusted forensic endpoint before detection and localization, so the endpoint still has access to the plaintext face, as it would in a conventional decrypt-then-infer pipeline. Given these results, it remains unclear which threat VeilTrace mitigates beyond that baseline.

We thank the reviewer for the comments. We would like to clarify the privacy objective and security boundaries of VeilTrace.

  1. Privacy protection against unauthorized access. VeilTrace aims to prevent direct exposure of the original face during transmission through recoverable image hiding. Detecting the presence of hidden content is different from recovering the original face. The additional key-dependent scrambling mechanism further restricts unauthorized reconstruction. However, we do not claim stronger confidentiality than properly implemented encryption.

  2. Steganalysis and concealment. We acknowledge that the near-perfect steganalysis results demonstrate limited statistical undetectability in the evaluated setting. Nevertheless, these results assess whether hidden content can be detected, rather than whether the original face can be reconstructed. We will distinguish these two security properties explicitly and revise the concealment claims accordingly.

  3. Trusted forensic endpoint. We agree that both VeilTrace and decrypt-then-infer require plaintext access at the trusted endpoint. VeilTrace does not aim to eliminate this exposure. Instead, it investigates whether sensitive images can be transmitted in a concealed representation and subsequently recovered while preserving forensic evidence for detection and localization. We will explicitly clarify this trust assumption and the distinction from encryption.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.