acceptodds
Under review as a conference paper at ICLR 2027

ViGuardQ: Quantized Video Deepfake Detection for Edge Devices

Abstract

The growing use of edge devices, e.g. smartphones, for video conferencing and facial biometric authentication creates an increasing demand for efficient video deepfake detection. However, the substantial computational and memory requirements of existing detectors hinder their deployment on resource-constrained devices. Although quantization offers an effective means of model compression, conventional methods do not explicitly account for the subtle spatial artifacts and temporal inconsistencies essential to video deepfake detection, which can be distorted by low-precision representations. To address this challenge, we propose ViGuardQ, a post-training quantization framework tailored to video deepfake detection on edge devices. First, Dual-Evidence-Constrained (DEC) Rotation mitigates forgery-evidence relevant quantization distortion by identifying a sensitive channel subspace using within-frame and cross-frame differential decision gradients, and optimizing orthogonal rotations to preserve more forgery-evidence sensitive information. Then Evidence-guided Consistent Response (ECR) Calibration refines the quantizer's consistency in response to evidence under aligned and opposite perturbations. They together improve the preservation of and responses to subtle forgery evidence and greatly improve quantization accuracy. Extensive experiments on six benchmark datasets and eleven deepfake detectors demonstrate that ViGuardQ outperforms state-of-the-art quantization baselines. Specifically, we achieve 83.59% AUC scores at a lower 13.4% compression rate, compared with 68.07% achieved by the strongest baseline, PTQ4VM, at a 14.95% compression rate on the DFD dataset. Real-world deployment on edge devices further demonstrates our practical effectiveness. Our code will be released upon acceptance.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Video Deepfake Detection, Quantization, Edge Deployment

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

ViGuardQ addresses post-training quantization of video deepfake detectors. Dual-Evidence-Constrained Rotation identifies spatial and temporal decision-sensitive feature directions and learns rotations and quantization scales while preserving the selected subspaces. Evidence-guided Consistent Response calibration subsequently matches the full-precision detector's responses to temporally aligned and opposing feature perturbations. The original detector weights remain frozen, with calibration performed on 256 videos. Experiments cover eleven detectors and six benchmarks, reporting improved AUC over quantization baselines and an Android deployment demonstration.

Critical strengths
  1. Preserving decision-sensitive directions and perturbation responses is a reasonable approach to limiting quantization damage to subtle forensic signals. The method operates on existing detectors without retraining their original weights, making the calibration setting practically relevant.

  2. The reported improvements extend across multiple architectures and datasets. The component ablations also provide useful evidence that DEC and ECR contribute to the reported gains.

Critical weaknesses
  1. The additional contribution beyond existing task-specific quantization is not sufficiently established. Preserving subtle forgery features during low-bit quantization is already central to Tao et al., "Unlocking the Potential of Lightweight Quantized Models for Deepfake Detection" (IJCAI 2025), which this submission cites. ViGuardQ's PTQ setting and temporal constraints provide plausible distinctions, but the evidence for the latter remains limited. The evaluation lists DeiT, ViT, Swin, and existing detectors without clearly specifying their temporal architectures, frame aggregation, or feature correspondence construction. Consequently, it is unclear whether the method preserves genuinely temporal detection behavior or improves calibration of predominantly framewise models. Appendix Table 6 shows benefits from adding TOP to TAP, but does not isolate temporal structure from additional perturbation coverage or calibration computation. A comparison against randomly paired or correspondence-shuffled perturbations under the same calibration budget would be substantially more informative.

  2. The reported accuracy-efficiency advantage is difficult to assess under the specified experimental conditions. The results do not clearly identify the weight and activation bit-widths, excluded full-precision layers, or whether quantization granularity is matched across methods. These choices directly affect both accuracy and resource use, so the large baseline gaps cannot yet be attributed confidently to the proposed design. Similarly, the Android figures of 90 ms per frame and 92 mW lack sufficient device, runtime, input, and measurement details, together with matched full-precision and quantized baseline measurements. DEC retains runtime rotations, but their contribution to the reported latency is not quantified. These omissions leave the central deployment claim insufficiently supported by the reported storage reductions alone.

Questions for authors
  1. Which evaluated detectors explicitly model temporal interactions, and how are matched feature locations obtained?

  2. Does the temporal benefit persist against random-pairing controls with matched perturbation counts and calibration budgets?

  3. What exact weight/activation precisions, quantization groups, and full-precision exceptions are used for each method? Are the comparisons matched on these settings?

6 comments
AuthorsOpenReview ↗

Response to Reviewer APAa's Weaknesses 1 and Question 1 & 2

Thank you for the detailed comments. Hope we address all your concerns and please drop us follow-up questions if there are any.

W1 & Q1:

W1.1 Difference between ViGuardQ and Tao et al. (IJCAI 2025).

ViGuardQ differs from Tao et al. (IJCAI 2025) in three main aspects:

1) Target Model difference. Tao et al. primarily focus on quantizing traditional CNN-based deepfake detectors, whereas our evaluation targets widely adopted more advanced Transformer-based architectures, including DeiT, ViT, and Swin. Given this, /experimental performance comparison is inapplicable/ .

2) Quantization Protocol difference. Tao et al. adopt quantization-aware /training/ (QAT), which jointly optimizes the detector and quantization components. In contrast, ViGuardQ employs post-training quantization (PTQ), preserving the pretrained detector weights and requiring only a small calibration dataset. /We substantially reduces training costs and enable deployment on resource-constrained edge devices without retraining. /

3) Method difference. (1)Tao et al. focus on preserving spatial forgery textures in individual images, whereas ViGuardQ additionally considers temporal inconsistencies across video frames. (2) Specifically, our DEC combines within-frame decision gradients with cross-frame gradient differences to preserve spatiotemporal forgery evidence, while ECR aligns FP and quantized decision responses under temporally aligned and opposed perturbations.

Q1.1 Clarification on Matched Feature Location.

We first perform face detection, and spatial alignment on adjacent video frames. Corresponding facial regions are then mapped to the feature grids at each selected layer, providing the matched locations used to compute cross-frame differential gradients.

AuthorsOpenReview ↗

Response to Reviewer APAa's Weaknesses 1 and Question 1 & 2 (Continued)

W1.2 Clarification of Temporal Infomation of Evaluated Detectors.

First, most advanced video deepfake detectors use image-based ViT backbones, encoding temporal information through input representations without modifying the underlying Transformer architecture [1–3]. Our evaluated detectors follow the same approach, incorporating temporal information through temporal encoding fusion while retaining the image-based ViT architectures. This motivates ViGuardQ to preserve discriminative spatial and temporal forgery evidence during quantization.

Second, to further demonstrate the generalizability of ViGuardQ, we evaluate two video-based transformers, ViViT[4] and FakeSTormer[5]: Both extend image-based vision transformer architectures with explicit temporal modeling mechanisms to capture cross-frame dependencies. Our results demonstrate that ViGuardQ can effectively quantize temporally enhanced detectors while maintaining competitive detection performance (Table 1 here).

Table 1. Evaluation on Video-based ViT Detectors (AUC %) on DFD dataset.

DatasetDetectorMetricFP (GT)BRECQAdaLogFIMA-QPTQ4VMViGuardQ
DFDViViTStorage Cost (MB) ↓355.6‾\underline{355.6}78.646.242.541.143.0
Compression Rate (%) ↓-22.113.012.011.612.1
Detection AUC (%) ↑85.6‾\underline{85.6}56.360.162.067.676.3
FakeSTormerStorage Cost (MB) ↓511.3‾\underline{511.3}113.661.960.058.260.4
Compression Rate (%) ↓-22.212.111.711.411.8
Detection AUC (%) ↑98.8‾\underline{98.8}59.272.373.179.489.6

Last, Our method explicitly uses temporal information during quantization process.

Theoretically, ViGuardQ operates on the input and output representations of detector blocks, enabling it to handle temporal information whether encoded at the feature level or explicitly modeled by the detector architecture. Specifically, DEC combines within-frame decision gradients and cross-frame gradient differences at corresponding facial locations (Eq. 6) to identify channels sensitive to spatial and temporal evidence. ECR uses temporally aligned and opposed feature perturbations to preserve the FP detector's signed changes in video-level decision margins. Together, these objectives enable ViGuardQ to preserve temporally relevant forensic evidence across different detector architectures.

Experimentally, we compare adjacent and nonadjacent frame pairings within the same facial tracks during DEC and ECR calibration to evaluate the contribution of temporal adjacency (Table 2 here). According to the results, using adjacent rather than nonadjacent pairs in both modules improves video-level AUC performance. Moreover, changing only ECR from adjacent to nonadjacent pairing reduces AUC value. These results indicate that temporally local frame correspondences are beneficial for preserving the FP detector's decision behavior during quantization.

Table 2: Ablation Study of Temporal Pairing in DEC and ECR (AUC %).

DFDDFDC
SettingDECECRDeiT-BViT-BSwin-BDeiT-BViT-BSwin-B
FP--88.893.793.275.677.877.4
A. Base--52.254.951.942.345.843.4
B1. DEC only✓ Adjacent-73.671.975.660.962.764.1
B2. DEC only✓ Nonadjacent-57.557.053.247.649.250.0
C1. Change both pairings✓ Nonadjacent✓ Nonadjacent68.265.470.352.150.456.7
C2. Change ECR pairing✓ Adjacent✓ Nonadjacent75.570.077.760.065.361.9
C3. DEC + ECR (full)✓ Adjacent✓ Adjacent83.584.985.566.669.868.5

[1]Han, Yue-Hua, et al. "Towards more general video-based deepfake detection through facial component guided adaptation for foundation model." 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025.

[2]Yan, Zhiyuan, et al. "Generalizing deepfake video detection with plug-and-play: Video-level blending and spatiotemporal adapter tuning." 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025.

[3]Yermakov, Andrii, et al. "Deepfake detection that generalizes across benchmarks." 2026 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). IEEE, 2026.

[4]Arnab, Anurag, et al. "Vivit: A video vision transformer." 2021 IEEE/CVF international conference on computer vision (ICCV). IEEE, 2021.

[5]Nguyen, Dat, et al. "Vulnerability-aware spatio-temporal learning for generalizable deepfake video detection." 2025 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2025.

AuthorsOpenReview ↗

Response to Reviewer APAa's Weaknesses 1 and Question 1 & 2 (Continued)

W1.3 & Q2 A comparison against randomly paired perturbations.

To test whether TOP benefits from matching the same facial region across adjacent frames, we kept TAP and the first-frame TOP location fixed and shifted the second-frame patch center from pt+1p_{t+1} to pt+1+dp_{t+1}+d, where dd is a random spatial offset drawn from a predefined neighborhood. The original pairing outperformed this control in all six settings across DFD and DFDC (Table 3 here). This result shows that where TOP is applied in the second frame matters, beyond simply adding another perturbation type.

Table 3: Effect of Temporal Pairing and Feature Correspondence (AUC%).

SettingECR perturbationsFrame pairing and feature correspondenceDFDDFDC
DeiT-BViT-BSwin-BDeiT-BViT-BSwin-B
DEC onlyNoneNo ECR calibration73.671.975.660.962.764.1
TAP onlyTAPMatched locations77.679.578.363.566.665.8
C2. Nonadjacent pairingTAP + TOPNonadjacent frames75.570.077.760.065.361.9
C1. Full ViGuardQTAP + TOPOriginal83.584.985.566.669.868.5
Random-offset TOPTAP + TOPRandom74.573.376.061.563.265.0
Reviewer_r4YvOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, the authors propose ViGuardQ, a post-training quantization framework for video deepfake detection on edge devices. Specifically, the method consists of two modules: DEC constructs feature subspaces based on intra-frame and inter-frame decision sensitivity, while ECR uses temporally aligned and opposed feature perturbations to align the quantized model’s responses to these changes with those of the full-precision model. The authors evaluate the framework on multiple datasets and 11 detection models and demonstrate its deployment on Android devices.

Critical strengths
  1. Deploying video deepfake detection models on resource-constrained devices is a meaningful problem.

  2. DEC and ECR have clear and complementary designs.

  3. The paper reports better results than the compared quantization baselines across multiple backbones, detection methods, and cross-dataset settings.

Critical weaknesses
  1. The experimental tables do not clearly specify the weight and activation bit widths used by each method, which layers remain in full precision, or whether the methods use the same calibration budget. Although Appendix D states that 256 calibration videos are used, key settings such as the subspace dimension, channel group size, perturbation magnitude, and number of optimization iterations are not provided. In addition, the relationship between the grouped activation quantization described in the main text and the “per-tensor” quantization stated in the appendix needs further explanation.

  2. The authors should report how much improvement DEC’s temporal sensitivity statistics provide over using spatial statistics alone, and whether ECR’s evidence-guided perturbation directions outperform random perturbations with the same energy.

  3. The experimental details for edge deployment remain limited. The authors report a latency of 90 ms per frame and power consumption of 92 mW, but do not sufficiently describe the device model, deployed model, inference backend, video input settings, or power measurement procedure. Comparisons with the full-precision model and quantization baselines on the same device are also missing.

Questions for authors

Please see weaknesses. Basiclly I hold a positive view on this paper, addressing the concerns can make this paper better supported.

1 comment
AuthorsOpenReview ↗

Response to Reviewer r4Yv's Weaknesses

Thank you for your positive assessment and detailed comments. We hope our response fully addresses your concerns, and please feel free to raise any further questions.

W1. Implementation and Setting Details.

All PTQ methods quantized all detector layers to W3A3 (3-bit weights and activations) and used the same 256 calibration videos from the FF++ training split. At each DEC interface, ViGuardQ identified a 32-dimensional decision-sensitive subspace within each group of 64 channels. The sensitive subspace and its 32-dimensional orthogonal complement each used a separate activation scale, giving two scales per group. ECR used a fixed perturbation magnitude of εℓ=0.5sℓ\varepsilon_\ell = 0.5 s_\ell for both TAP and TOP at every selected layer. With eight clips per update, ViGuardQ performed 128 DEC updates followed by 64 ECR updates, for a total of 192 optimization steps per detector. The “per-tensor” description in Appendix D refers to quantization within each subgroup rather than a single scale shared across the entire layer. We will revise this terminology to “group-wise activation quantization” to avoid ambiguity.

W2. More ablations on DEC’s temporal sensitivity and ECR’s evidence-guided perturbation.

DEC uses cross-frame sensitivity to identify which feature directions to preserve during quantization, while ECR uses the full-precision detector’s sensitivity to guide perturbations within that subspace. To test these choices separately, we compared joint spatial–temporal DEC with spatial-only DEC while keeping ECR disabled, then compared full ECR with random directions while keeping DEC, the TAP/TOP patterns, and perturbation energy fixed (Table 1 here). Across three backbones on DFD and DFDC, adding temporal statistics improved mean AUC by /10.67/ %, and including evidence-guided directions improved mean AUC by /6.53/ % over random directions alone.

Table 1: Ablation of DEC Sensitivity and ECR Perturbation Directions (AUC %).

ModuleAblation settingDFDDFDC
DeiT-BViT-BSwin-BDeiT-BViT-BSwin-B
Base-52.254.951.942.345.843.4
DEC onlyD1. Spatial sensitivity (α = 1)65.367.063.550.050.148.9
DEC onlyD2. Temporal sensitivity (α = 0)67.268.165.551.352.855.5
DEC onlyD3. Joint spatial and temporal73.671.975.660.962.764.1
DEC + ECRE1. Random directions-only77.074.378.561.163.765.0
DEC + ECRE2. Proposed ECR83.584.985.566.669.868.5

W3. Deployment details and comparisons.

The Details of Deployment. The Android demo was conducted on a smartphone (model PKB110) equipped with a MediaTek Dimensity 9400 SoC and 16 GB of physical RAM (with 12 GB of virtual RAM expansion).

Deployment comparison on the same Android device. According to the Table 2 (here), ViGuardQ runs at 11 FPS (90 ms/frame) and draws 85% less power than FP32 (92 vs. 613 mW). Compared with the PTQ baseline (BRECQ), ViGuardQ achieves a favorable power–accuracy trade-off: it raises AUC from 60% to 85%.

Table 2: On-device Performance Comparison of FP32, BRECQ (PTQ baseline), and ViGuardQ.

MethodLatency (ms) ↓FPS ↑Runtime Memory (MB) ↓Peak Memory (MB) ↓Power (mW) ↓AUC (%) ↑
FP3234031200165061390.0
BRECQ80123204709060.0
ViGuardQ90113404859285.0
Reviewer_SCUvOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents ViGuardQ, a post-training quantization (PTQ) framework for video deepfake detection on edge devices. It introduces two components. DEC Rotation uses intra-frame and inter-frame decision gradients to identify sensitive channel subspaces and optimize rotations. ECR Calibration matches the decision-margin changes of full-precision and quantized models under temporally aligned and opposed perturbations.

The authors evaluate ViGuardQ on six deepfake benchmarks and 11 detectors. The results show large AUC improvements over existing PTQ baselines, along with reduced average model storage. The paper also includes module ablations and an Android deployment experiment.

Critical strengths
  1. The paper addresses an important challenge in deploying deepfake detectors on resource-constrained devices. I like that the proposed method considers the detector's decision behavior rather than focusing only on generic feature reconstruction errors.

  2. DEC and ECR address two different aspects of quantization distortion. The use of decision-sensitive subspaces and signed response matching is interesting. I also appreciate the mathematical explanations provided in the appendix.

  3. It provides an encouraging experimental results. The reported AUC improvements are substantial across different detectors and datasets. Tables 3 and 6 provide useful ablations showing the contributions of DEC, ECR, TAP, and TOP. The Android deployment experiment is also a useful first step toward practical evaluation.

Critical weaknesses
  1. The video-specific motivation is not sufficiently supported by the experimental setting. I am not fully convinced that the experiments directly test the proposed temporal motivation. Several evaluated detectors are originally image-based, but the paper does not clearly explain how they process multiple frames or produce video-level predictions. This makes it difficult for me to understand whether the temporal objectives preserve information that these detectors actually use.

  2. The contributions of the core mechanisms are not fully isolated. The ablations show that DEC and ECR improve AUC, but I am less clear about why the specific designs are necessary. For DEC, spatial-only and temporal-only sensitivity comparisons are missing. For ECR, I would like to see whether the improvement comes from response matching or simply from additional calibration. Direct measurements of response consistency would also help verify the claimed mechanism.

  3. The quantization settings and storage comparisons need clarification. I could not find the exact weight and activation bit-widths, quantized layers, and baseline configurations. This matters because Table 1 reports quite different storage ratios. For example, DeiT-Base is reduced from 330.2 MB to 33.9 MB (about 10.3%), below the nominal 12.5% ratio for FP32-to-INT4 weight quantization. There may be a reasonable explanation, but I would like to understand how these numbers are obtained.

  4. The edge deployment evaluation is still limited. Section 4.2 reports 90 ms/frame and 92 mW on Android, which is encouraging. However, the device, inference backend, end-to-end latency, and peak memory are not sufficiently reported. Without comparisons against full-precision and other quantized models on the same device, I find it difficult to assess the actual deployment benefits. The offline calibration cost is also not reported.

  5. The positioning against some related methods could be improved. I would like to see a clearer discussion of how DEC differs from existing learned-rotation quantization methods, such as SpinQuant [1], and how ECR relates to earlier work on preserving teacher responses, such as Jacobian matching [2]. These works address different settings, but their technical connections seem relevant. I also noticed that Table 2 categorizes GAC-FAS as Neural Architecture Search, although the cited work proposes gradient alignment for face anti-spoofing. This makes part of the related-work comparison difficult to follow.

[1] Z. Liu, C. Zhao, I. Fedorov, B. Soran, D. Choudhary, R. Krishnamoorthi, V. Chandra, Y. Tian, and T. Blankevoort, "SpinQuant: LLM Quantization with Learned Rotations," in International Conference on Learning Representations (ICLR), 2025

[2] S. Srinivas and F. Fleuret, "Knowledge Transfer with Jacobian Matching," in Proceedings of the 35th International Conference on Machine Learning (ICML), 2018

Questions for authors

Q1. Could the authors clarify how the evaluated detectors process video frames and generate video-level predictions? Do any of them explicitly model temporal dependencies?

Q2. Could the authors provide spatial-only and temporal-only DEC ablations, and compare ECR with standard calibration under the same optimization budget? Reporting FP–quantized response errors on held-out data would also be helpful.

Q3. Could the authors specify the W/A bit-widths, quantized layers, and storage accounting for each method? In particular, how are the storage ratios below the nominal INT4 ratio achieved, and how is PTQ4VM adapted to the evaluated detectors?

Q4. Could the authors report the Android hardware and runtime settings, end-to-end FPS, peak memory, and latency/power comparisons against FP and PTQ baselines on the same device? What is the full DEC/ECR calibration cost?

Q5. Could the authors explain the differences from related rotation-based PTQ and response-matching methods [1,2], and check the technical categories used in Table 2?

9 comments
AuthorsOpenReview ↗

Response to Reviewer SCUv’s Weakness 1 and Question 1

Thank you for the detailed comments. Hope we address all your concerns and please drop us follow-up questions if there are any.

W1 & Q1. Clarification and additional experiments of evaluated detectiors' capability in handling temporal information.

(1) Clarification of evaluation of Image-based ViT detector. Generally, most advanced video deepfake detectors use image-based ViT backbones, encoding temporal information through input representations without modifying the underlying Transformer architecture [1–3]. Our evaluated detectors follow the same approach, incorporating temporal information through temporal encoding fusion while retaining the image-based ViT architectures. This motivates ViGuardQ to preserve discriminative spatial and temporal forgery evidence during quantization.

(2) More evaluation on Video-Adapted ViT detector. In Table.1 (here), we further evaluate our method on two additional ViT-based detectors with model architecutre adapted for video processing, ViViT [4] and FakeSTormer [5]. Although they both build on image-based ViT backbones. Our improvement on accuracy over other quantization baselines are still significant under comprarable compression rate.

Table 1. Evaluation on Video-based ViT Detectors (AUC %) on DFD dataset.

DatasetDetectorMetricFP (GT)BRECQAdaLogFIMA-QPTQ4VMViGuardQ
DFDViViTStorage Cost (MB) ↓355.6‾\underline{355.6}78.646.242.541.143.0
Compression Rate (%) ↓-22.113.012.011.612.1
Detection AUC (%) ↑85.6‾\underline{85.6}56.360.162.067.676.3
FakeSTormerStorage Cost (MB) ↓511.3‾\underline{511.3}113.661.960.058.260.4
Compression Rate (%) ↓-22.212.111.711.411.8
Detection AUC (%) ↑98.8‾\underline{98.8}59.272.373.179.489.6

(3) More detailed explanantion on how we handle temporal clues.

Theoretically, ViGuardQ operates on the input and output representations of detector blocks, enabling it to handle temporal information whether encoded at the feature level or explicitly modeled by the detector architecture. Specifically, DEC combines within-frame gradients and cross-frame differences to identify spatially and temporally sensitive channels, while ECR uses aligned and opposed perturbations to preserve FP decision responses. Together, they enable ViGuardQ to retain temporal forensic evidence across detector architectures.

Experimentally, (1) we compare adjacent and nonadjacent frame pairings within the same facial tracks during DEC and ECR calibration to evaluate the contribution of temporal adjacency in Table 2 (here). According to the results, using adjacent rather than nonadjacent pairs in both modules improves video-level AUC performance. Moreover, changing only ECR from adjacent to nonadjacent pairing reduces AUC value. These results indicate that temporally local frame correspondences are beneficial for preserving the FP detector's decision behavior during quantization.

Table 2: Ablation Study of Temporal Pairing in DEC and ECR (AUC %).

DFDDFDC
SettingDECECRDeiT-BViT-BSwin-BDeiT-BViT-BSwin-B
FP--88.893.793.275.677.877.4
A. Base--52.254.951.942.345.843.4
B1. DEC only✓ Adjacent-73.671.975.660.962.764.1
B2. DEC only✓ Nonadjacent-57.557.053.247.649.250.0
C1. Change both pairings✓ Nonadjacent✓ Nonadjacent68.265.470.352.150.456.7
C2. Change ECR pairing✓ Adjacent✓ Nonadjacent75.570.077.760.065.361.9
C3. DEC + ECR (full)✓ Adjacent✓ Adjacent83.584.985.566.669.868.5

[1]Towards more general video-based deepfake detection through facial component guided adaptation for foundation model. 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025.

[2]Generalizing deepfake video detection with plug-and-play: Video-level blending and spatiotemporal adapter tuning. 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025.

[3]Deepfake detection that generalizes across benchmarks. 2026 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). IEEE, 2026.

[4]Vivit: A video vision transformer. 2021 IEEE/CVF international conference on computer vision (ICCV). IEEE, 2021.

[5]Vulnerability-aware spatio-temporal learning for generalizable deepfake video detection. 2025 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2025.

AuthorsOpenReview ↗

Response to Reviewer SCUv’s Weakness 1 and Question 1 (Continued)

(4) More Ablation study on Effect of Temporal Pairing and Feature Correspondence. To test whether TOP benefits from matching the same facial region across adjacent frames, we kept TAP and the first-frame TOP location fixed and shifted the second-frame patch center from pt+1p_{t+1} to pt+1+dp_{t+1}+d, where dd is a random spatial offset drawn from a predefined neighborhood. The original pairing outperformed this control in all six settings across DFD and DFDC (Table 3 here). This result shows that where TOP is applied in the second frame matters, beyond simply adding another perturbation type.

Table 3: Effect of Temporal Pairing and Feature Correspondence (AUC%).

SettingECR perturbationsFrame pairing and feature correspondenceDFDDFDC
DeiT-BViT-BSwin-BDeiT-BViT-BSwin-B
DEC onlyNoneNo ECR calibration73.671.975.660.962.764.1
TAP onlyTAPMatched locations77.679.578.363.566.665.8
C2. Nonadjacent pairingTAP + TOPNonadjacent frames75.570.077.760.065.361.9
C1. Full ViGuardQTAP + TOPOriginal83.584.985.566.669.868.5
Random-offset TOPTAP + TOPRandom74.573.376.061.563.265.0
AuthorsOpenReview ↗

Response to Reviewer SCUv’s Weakness 2 and Question 2

W2 & Q2: Additional ablations across three backbones (DeiT-B, ViT-B, and Swin-B).

(1) To isolate each module’s contribution, we varied DEC’s sensitivity sources (Table 3a here). DEC achieved higher AUC by combining within-frame and cross-frame sensitivities than by using either alone

(2) We also varied ECR’s calibration target (Table 3b here). ECR achieved higher AUC by matching the FP detector’s signed margin changes under TAP and TOP than by continuing feature reconstruction or matching the perturbed margins directly.

(3) In addition, we additionally measure normalized FP-quantized response errors to directly evaluate the claimed mechanism (Table 3a & 3b here).

Table 3a: Ablation on DEC Sensitivity. Base: PTQ baseline. D1: Spatial-guided sensitivity (α = 1). D2: Temporal-guided sensitivity (α = 0). D3: Joint spatial and temporal sensitivity (Our DEC only).

SettingDFDDFDC
DeiT-B AUC ↑ViT-B AUC ↑Swin-B AUC ↑Response Error ↓DeiT-B AUC ↑ViT-B AUC ↑Swin-B AUC ↑Response Error ↓
Base52.254.951.912.542.345.843.414.9
D165.367.063.510.150.050.148.911.3
D267.268.165.510.751.352.855.510.9
D3 (DEC)73.671.975.68.260.962.764.19.8

Table 3b: Ablation on ECR calibration. E1: Additional calibration with traditional reconstruction. E2: Absolute-margin matching on TAP/TOP. E3: Full ViGuardQ (DEC + ECR).

SettingDFDDFDC
DeiT-B AUC ↑ViT-B AUC ↑Swin-B AUC ↑Response Error ↓DeiT-B AUC ↑ViT-B AUC ↑Swin-B AUC ↑Response Error ↓
D3 (DEC)73.671.975.68.260.962.764.19.8
E175.573.778.37.562.163.064.59.6
E278.277.981.27.163.464.566.68.0
E3 (Ours)83.584.985.53.366.669.868.54.5

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response: Summary of New Results and Clarifications

We sincerely thank all reviewers for their constructive feedback and valuable suggestions. In response to their concerns, we provide additional experiments and technical clarifications to further strengthen our contributions:

  1. Temporal Information Excavation: We clarify how image-based ViT detectors process multi-frame video evidence without modifying their original Transformer backbones. We further extend our evaluation to ViViT and FakeSTormer, which explicitly model temporal dependencies, demonstrating ViGuardQ's applicability across different detector architectures.

  2. Additional Experimental Evidence: We evaluate two video-based ViT detectors, ViViT and FakeSTormer, achieving 8.7 and 10.2 AUC improvements over PTQ4VM on DFD, respectively. Additional ablations demonstrate the effectiveness of temporal sensitivity (+10.67%), evidence-guided perturbations (+6.53%), temporal adjacency (+8.07%), and feature correspondence (+7.55%).

  3. Quantization Fairness and Edge Deployment: We clarify the consistent W3A3 quantization protocol, using 256 FF++ calibration videos, and explain storage variations caused by weight packing and quantization scales. On Android, ViGuardQ achieves 90 ms/frame (11 FPS), 485 MB peak memory, and 92 mW power, demonstrating its practical accuracy–efficiency trade-off.

To address the reviewers' technical and experimental concerns, we summarize the additional experiments and provide the following index for convenient cross-reference:

  • Table A (Video-based ViT Quantization): Comparisons of ViGuardQ with BRECQ, AdaLog, FIMA-Q, and PTQ4VM on ViViT and FakeSTormer (SCUv-W1; APAa-W1).
  • Table B (Temporal Pairing Ablation): Adjacent versus nonadjacent frame pairing in DEC and ECR (SCUv-W1; APAa-W1).
  • Table C (Feature Correspondence Ablation): TAP/TOP calibration with original, nonadjacent, and random-offset correspondences (SCUv-W1; APAa-W1).
  • Table D (DEC Sensitivity Ablation): Spatial-only, temporal-only, and joint decision sensitivity, with response-consistency evaluation (SCUv-W2; r4Yv-W2).
  • Table E (ECR Calibration Ablation): Conventional reconstruction, absolute-margin matching, and signed-response matching (SCUv-W2).
  • Table F (Evidence-guided Perturbation Ablation): Comparison between random and evidence-guided perturbation directions under matched perturbation energy (r4Yv-W2).
  • Table G (Android Deployment): On-device latency, FPS, memory, power, and detection AUC comparisons with FP32 and PTQ baselines (SCUv-W4; r4Yv-W3; APAa-W2).

We incorporated these experimental results, methodological clarifications, and related-work corrections into the revised manuscript. We hope these additional results and explanations adequately address the reviewers' concerns and further demonstrate the effectiveness and practicality of ViGuardQ.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.