Back

Robust Image Forgery Detection over Online Social Network Shared Images

Haiwei WuJiantao ZhouJinyu TianJun Liu

article2022CVPR102 citations

Proposes a training framework that decouples social network transmission artifacts into predictable and adversarial unseen noise, enabling forensic detectors to reliably identify manipulated images degraded by online compression and resizing.

Listen

Online social networks have become the primary channels for distributing manipulated digital images, impacting legal proceedings, news integrity, and public trust. However, the automated, lossy operations used by these platforms—such as image resizing, enhancement filtering, and compression—frequently erase or distort the subtle digital artifacts needed to identify tampered content. Standard forensic detectors suffer severe performance degradation once an image is uploaded and shared across these channels.

The article develops and evaluates a robust training framework designed to maintain high-accuracy image forgery detection even when images pass through lossy online platforms.

To address this challenge without relying on restricted or opaque platform algorithms, the researchers decoupled platform-induced interference into two components: predictable noise and unseen noise. Predictable noise, such as standard JPEG compression, was modeled using a neural network paired with a specialized mathematical layer that simulates compression while allowing end-to-end model training. Unseen noise—capturing unknown filters, adaptive processing, or variations across platforms—was modeled using adversarial noise techniques, which introduce subtle worst-case perturbations to stress-test the detector. The framework was trained using platform data from Facebook and tested across 5,232 images sourced from four standard forensic datasets transmitted across Facebook, Weibo, and WeChat.

Key findings show that the proposed framework substantially outperforms existing forensic models under real-world conditions. While competitor detectors experienced severe performance drops—with overlap detection metrics falling by roughly 9% to 11% following social network transmission—the proposed method experienced minimal loss, dropping by only 0.9% on Facebook, 2.0% on Weibo, and 4.5% on WeChat. On average, across all platforms, the system achieved an area under the curve score of approximately 0.83 to 0.85 and an intersection-over-union score around 0.36 to 0.40, roughly double the accuracy of existing state-of-the-art tools. Ablation analysis confirmed that combining both predictable and unseen noise models yielded the strongest gains, improving balanced accuracy scores by over 15% on transmitted images. Furthermore, the model generalized effectively to Weibo and WeChat despite being trained exclusively on Facebook data, and it demonstrated superior resilience against common post-processing operations like blurring, cropping, and general noise addition.

These results demonstrate that forensic detectors can be hardened against real-world transmission channels without requiring inside knowledge of proprietary platform pipelines. For organizations managing digital trust, compliance, or misinformation risks, adopting noise-resilient training approaches significantly reduces the operational risk of false negatives when vetting shared media.

Decision-makers should consider integrating dual-noise training principles into their automated content moderation and forensic pipelines. Because the authors have made their benchmark dataset and codebase publicly available, technical teams can readily benchmark internal detectors. Next steps should focus on pilot implementations and evaluating performance against emerging generative media threats.

The findings are supported by consistent results across multiple standard datasets and platforms. A minor limitation is that the model exhibited slightly higher degradation on platforms with stricter compression policies, such as WeChat. Nevertheless, the methodology provides a highly reliable foundation for practical forensic detection across modern communication networks.

University of Macauimage forgeryimage tamperingonline social networksrobust image forgery detectionimage forensicsimage manipulation detection
Wu et al (2022).pdf
Cover for Robust Image Forgery Detection over Online Social Network Shared Images

Abstract

The increasing abuse of image editing softwares, such as Photoshop and Meitu, causes the authenticity of digital images questionable. Meanwhile, the widespread availability of online social networks (OSNs) makes them the dominant channels for transmitting forged images to report fake news, propagate rumors, etc. Unfortunately, various lossy operations adopted by OSNs, e.g., compression and resizing, impose great challenges for implementing the robust image forgery detection. To fight against the OSN-shared forgeries, in this work, a novel robust training scheme is proposed. We first conduct a thorough analysis of the noise introduced by OSNs, and decouple it into two parts, i.e., predictable noise and unseen noise, which are modelled separately. The former simulates the noise introduced by the disclosed (known) operations of OSNs, while the latter is designed to not only complete the previous one, but also take into account the defects of the detector itself. We then incorporate the modelled noise into a robust training framework, significantly improving the robustness of the image forgery detector. Extensive experimental results are presented to validate the superiority of the proposed scheme compared with several state-of-the-art competitors. Finally, to promote the future development of the image forgery detection, we build a public forgeries dataset based on four existing datasets and three most popular OSNs. The designed detector recently won the top ranking in a certificate forgery detection competition^1. The source code and dataset are available at https://github.com/HighwayWu/ImageForensicsOSN.

Topics

image forgeryimage tamperingonline social networkshigh resolution image

Keywords

robust image forgery detectionimage forensicsimage manipulation detectionimage copy-move forgery detectionimage splicingonline social networkscamera model fingerprintforensic similaritydifferentiable JPEGJPEG compression

Table of Contents

  • 1. Introduction
  • 2. Related works
  • 2.1. Image forgery detection
  • 2.2. Online Social Network (OSN)
  • 3. Robust image forgery detection against transmission over OSNs
  • 3.1. Modeling the distribution P(τ)
  • 3.2. Modeling the conditional distribution P(ξ | τ)
  • 4. Experimental results
  • 4.1. Experimental setup
  • 4.2. Quantitative comparisons
  • 4.3. Qualitative comparisons
  • 4.4. Ablation studies
  • 4.5. Some further robustness evaluations
  • 5. Conclusions
  • References

Knowls

  1. Knowl 1 — Compound-noise robust training objective

    model/method

    The proposed detector-training framework models degradation from online social networks as the sum of predictable noise and unseen noise. For a forged RGB image xi∈RH×W×3x_i\in\mathbb{R}^{H\times W\times 3} with binary forgery mask yi∈{0,1}H×W×1y_i\in\{0,1\}^{H\times W\times 1}, the compound perturbation is δ=τ+ξ\delta=\tau+\xi, where τ\tau is predictable OSN noise and ξ\xi is unseen noise. The detector fθf_\theta outputs a pixelwise forgery-probability map, and its parameters are trained by minimizing the expected binary cross-entropy loss:

    min⁡θ1N∑i=1NEτ∼P(τ)[Eξ∼P(ξ∣τ)[Lb(fθ(xi+τ+ξ),yi)]].\min_\theta \frac{1}{N}\sum_{i=1}^{N}\mathbb{E}_{\tau\sim P(\tau)}\left[\mathbb{E}_{\xi\sim P(\xi\mid\tau)}\left[\mathcal{L}_b\big(f_\theta(x_i+\tau+\xi),y_i\big)\right]\right].

    The training forgeries are synthesized from two pristine color images p1,p2∈RH×W×3p_1,p_2\in\mathbb{R}^{H\times W\times 3} and a mask yy using x=p1⊙(1−y)+p2⊙yx=p_1\odot(1-y)+p_2\odot y, where ⊙\odot denotes elementwise multiplication and the mask value 11 identifies the inserted region. The nested expectation permits the two noise components to be statistically dependent rather than assuming independent OSN degradations.

  2. Knowl 2 — Differentiable surrogate for predictable OSN noise

    model/method

    Predictable OSN noise is generated with a surrogate image-to-image network gϕ:Rd→Rdg_\phi:\mathbb{R}^{d}\rightarrow\mathbb{R}^{d} trained from pairs (xi,OSN⁡(xi))(x_i,\operatorname{OSN}(x_i)), where d=H×W×3d=H\times W\times 3 and OSN⁡(⋅)\operatorname{OSN}(\cdot) denotes the complete processing pipeline of a chosen social network. The surrogate uses a U-Net architecture and residual learning: it predicts a residual that is added to the input before JPEG processing. Its training objective is

    min⁡ϕLr(Jq(xi+gϕ(xi)),OSN⁡(xi)),Lr(a,b)=∥a−b∥2,\min_\phi \mathcal{L}_r\left(J_q\big(x_i+g_\phi(x_i)\big),\operatorname{OSN}(x_i)\right),\qquad \mathcal{L}_r(a,b)=\lVert a-b\rVert_2,

    where JqJ_q is a differentiable JPEG-compression layer with quality factor qq. The JPEG quantization step replaces ordinary nondifferentiable rounding with the smooth approximation ⌊z⌋+(z−⌊z⌋)3\lfloor z\rfloor+(z-\lfloor z\rfloor)^3 for scalar coefficient zz. After optimizing gϕg_\phi to gϕ∗g_{\phi^*}, predictable noise samples are produced as

    τi(q)=Jq(xi+gϕ∗(xi))−xi.\tau_i(q)=J_q\big(x_i+g_{\phi^*}(x_i)\big)-x_i.

    During training, qq is sampled uniformly from the interval [71,95][71,95], matching the quality-factor range observed for Facebook. The surrogate is intended to reproduce signal-dependent effects such as resizing, enhancement filtering, and JPEG compression without repeatedly uploading and downloading every training image.

  3. Knowl 3 — Adversarial model of unseen OSN noise

    model/method

    Unseen noise represents unknown OSN operations, inaccurate parameter estimates, differences between the training and testing platforms, and degradation sources not captured by the predictable-noise surrogate. Instead of modeling these unknown operations directly from image signals, the method models only perturbations that damage the forgery detector. For image xix_i, predictable noise τi\tau_i, target mask yiy_i, detector fθf_\theta, and binary-cross-entropy loss Lb\mathcal{L}_b, an input-specific unseen perturbation is chosen along the signed input gradient:

    ξi=S(∇xiLb(fθ(xi+τi),yi)),\xi_i=\mathcal{S}\left(\nabla_{x_i}\mathcal{L}_b\big(f_\theta(x_i+\tau_i),y_i\big)\right),

    where S\mathcal{S} applies the elementwise sign function. To make the perturbation less specific to one training image, the method estimates a global gradient direction from previously processed samples. With ξ0=0\xi_0=0, the direction for step t+1t+1 is based on

    ξt+1=1t∑i=0tS(∇xiLb(fθ(xi+τi+ξi),yi)).\xi_{t+1}=\frac{1}{t}\sum_{i=0}^{t}\mathcal{S}\left(\nabla_{x_i}\mathcal{L}_b\big(f_\theta(x_i+\tau_i+\xi_i),y_i\big)\right).

    The global direction is then modeled stochastically as a conditional Gaussian:

    ξt+1∣τ∼N(ut+1,σ2I),ut+1=ϵ1t∑i=0tS(∇xiLb(fθ(xi+τi+ξi),yi)),\xi_{t+1}\mid\tau\sim\mathcal{N}(u_{t+1},\sigma^2 I),\qquad u_{t+1}=\epsilon\frac{1}{t}\sum_{i=0}^{t}\mathcal{S}\left(\nabla_{x_i}\mathcal{L}_b\big(f_\theta(x_i+\tau_i+\xi_i),y_i\big)\right),

    where II is the identity matrix, σ\sigma controls random variation around the global direction, and ϵ\epsilon constrains perturbation magnitude. This construction focuses training on small, detector-degrading perturbations that serve as a proxy for otherwise unobservable OSN noise.

  4. Knowl 4 — Monte Carlo robust-training procedure

    algorithm

    The complete training procedure first learns the predictable-noise surrogate and then trains the forgery detector with Monte Carlo samples of both noise components. Its inputs are an OSN-paired dataset D1D_1, a forged-image-and-mask dataset D2D_2, epoch counts N1,N2N_1,N_2, learning rates lϕ,lθl_\phi,l_\theta, predictable-noise sample count mm, unseen-noise sample count hh, Gaussian scale σ\sigma, and perturbation scale ϵ\epsilon. It returns the trained detector fθ∗f_{\theta^*}.

    Input: OSN-paired dataset D1; forged-image dataset D2; epochs N1 and N2; learning rates lφ and lθ; sample counts m and h; σ and ε
    Output: Trained forgery detector fθ*
    Randomly initialize φ and θ
    for epoch = 1 to N1 do
        for minibatch x from D1 do
            Sample JPEG quality factor q uniformly from [71, 95]
            Compute gφ-gradient of Lr(Jq(x + gφ(x)), OSN(x))
            Update φ = φ - lφ times gφ
        end
    end
    Set φ* = φ and initialize u = 0
    for epoch = 1 to N2 do
        for minibatch (x, y) from D2 do
            Set accumulated loss L = 0
            for j = 1 to m do
                Sample qj uniformly from [71, 95]
                Compute predictable noise τj = Jqj(x + gφ*(x)) - x
                Sample ξ1 through ξh from the Gaussian N(u, σ²I)
                Add sum over k of Lb(fθ(x + τj + ξk), y) to L
            end
            Compute detector gradient gθ = ∇θL and input gradient gx = ∇xL
            Update θ = θ - lθ times gθ
            Update u = u + ε times S(gx)
        end
    end
    Return fθ* = fθ

    The corresponding Monte Carlo training loss approximates the nested expectation with mm samples of predictable noise and hh samples of unseen noise for each predictable-noise sample:

    LMC(θ)=∑i=1N∑j=1m∑k=1hLb(fθ(xi+τij+ξijk),yi).\mathcal{L}_{\mathrm{MC}}(\theta)=\sum_{i=1}^{N}\sum_{j=1}^{m}\sum_{k=1}^{h}\mathcal{L}_b\big(f_\theta(x_i+\tau_{ij}+\xi_{ijk}),y_i\big).

    The paper specifies the JPEG-quality-factor range but does not provide numerical values for every other implementation hyperparameter in the supplied text.

  5. Knowl 5 — Pixel-level detector and training configuration

    experimental setup

    The forgery detector is a pixel-level mapping fθ:RH×W×3→RH×W×1f_\theta:\mathbb{R}^{H\times W\times 3}\rightarrow\mathbb{R}^{H\times W\times 1} that predicts a binary forgery mask. The baseline architecture is a U-Net augmented with spatial-channel Squeeze-and-Excitation blocks, called SE-U-Net. The predictable-noise network is trained using WEI, which contains more than 1,300 original images and their Facebook-transmitted versions. The forgery detector is trained from pristine images in Dresden, with inserted objects taken from MS-COCO to create spliced images and corresponding masks. The WEI and synthesized-forgery datasets are each split randomly into training and validation subsets in a 9:1 ratio. Comparisons use MT-Net, NoiPri, ForSim, DFCN, and an SE-U-Net baseline without the proposed robust training.

  6. Knowl 6 — Public OSN-transmitted forgery benchmark

    data/table

    The paper constructs a public benchmark by taking four forgery datasets—DSO, Columbia, NIST, and CASIA—and manually uploading and downloading their images through Facebook, Weibo, and WeChat. The resulting collection contains 5,232 OSN-transmitted forgeries with corresponding pixel masks. It evaluates both the original images and their transmitted versions, allowing robustness to be measured separately for each OSN. The predictable-noise surrogate is trained only with Facebook examples; Weibo and WeChat data are reserved for testing cross-platform generalization. The benchmark and source code are released publicly, and the dataset is intended to support pixel-level forgery localization under realistic social-network processing.

  7. Knowl 7 — Average pixel-level comparison across OSN conditions

    data/table

    The main quantitative evaluation averages AUC, F1, and IoU over the DSO, Columbia, NIST, and CASIA test sets; larger values are better. It compares five existing or baseline detectors with the proposed robustly trained detector under no OSN transmission and under Facebook, Weibo, and WeChat transmission.

    Could not parse LaTeX table

    The proposed detector has the highest average AUC, F1, and IoU in every transmission condition. Unlike all competing methods, it preserves nearly the same performance after OSN processing: its average IoU is .409 without transmission, .400 after Facebook, .380 after Weibo, and .364 after WeChat.

  8. Knowl 8 — Cross-platform robustness from Facebook-only noise training

    empirical result

    OSN transmission substantially damages conventional forgery localization. For example, the paper reports that MT-Net's IoU decreases by 10.1%, 11.1%, and 9.4% after Facebook, Weibo, and WeChat transmission, respectively, relative to its no-transmission performance. The proposed detector instead has reported IoU reductions of only 0.9%, 2.0%, and 4.5% for the same platforms. Weibo and WeChat are more damaging than Facebook because their image-processing pipelines apply more stringent compression and consequently remove more forensic evidence. Despite using only Facebook-transmitted images to train the predictable-noise surrogate, the proposed detector generalizes to both unseen OSN platforms and remains the best-performing method on their transmitted test sets.

  9. Knowl 9 — Ablation of predictable and unseen noise

    data/table

    Ablation experiments isolate the contributions of predictable noise τ\tau and unseen noise ξ\xi using the SE-U-Net detector, with and without Facebook transmission. The values are AUC, F1, and IoU; parenthesized values are gains over the corresponding baseline.

    Could not parse LaTeX table

    Predictable noise alone gives modest improvements, while adversarial unseen noise gives larger gains. Combining both components yields the strongest robustness, including a 15.7-percentage-point F1 gain over the SE-U-Net baseline under Facebook transmission. Applying the same compound-noise training to a DPN detector also improves performance, indicating that the training strategy is not restricted to the SE-U-Net architecture.

  10. Knowl 10 — Robustness to non-OSN post-processing

    empirical result

    The proposed detector is additionally evaluated on the Columbia forgery set after applying cropping, resizing, blurring, additive noise, or standalone JPEG compression. A unified parameter pp controls the intensity of each operation, with p=0p=0 denoting the unprocessed images. As pp increases, the competing detectors MT-Net, NoiPri, ForSim, and DFCN do not maintain consistent localization performance. The proposed robustly trained detector generalizes more consistently across all five degradation types, showing that the predictable-plus-unseen noise training strategy improves robustness beyond the specific OSN pipelines used to construct the training data.

Coverage note — The paper's qualitative visualization examples were not made into a separate knowl because they illustrate the quantitative robustness findings rather than add an independent result.

References

  1. 1.Nist nimble 2016 datasets. https://www.nist.gov/itl/iad/mig/nimble-challenge-2017-evaluation/. 2, 6, 7
  2. 2.I. Amerini, T. Uricchio, L. Ballan, and R. Caldelli. Localization of jpeg double compression through multi-domain convolutional neural networks. In Proc. IEEE Conf. Comput. Vis. Pattern Recogn. Workshop, 2017. 2
  3. 3.Q. Bammey, R. Gioi, and J. Morel. An adaptive neural network for unsupervised mosaic consistency analysis in image forensics. In Proc. IEEE Conf. Comput. Vis. Pattern Recogn., pages 14194–14204, 2020. 2
  4. 4.B. Bayar and M. C. Stamm. Constrained convolutional neural networks: a new approach towards general purpose image manipulation detection. IEEE Trans. Inf. Forensics and Security, 13(11):2691–2706, 2018. 1, 2
  5. 5.L. Bondi, S. Lameri, D. Guera, P. Bestagini, E. J. Delp, and S. Tubaro. Tampering detection and localization through clustering of camera-based cnn features. In Proc. IEEE Conf. Comput. Vis. Pattern Recogn. Workshops, pages 1855–1864, 2017. 1, 2
  6. 6.T. Carvalho, C. Riess, E. Angelopoulou, H. Pedrini, and A. Rezende Rocha. Exposing digital image forgeries by illumination color classification. IEEE Trans. Inf. Forensics and Security, 8(7):1182–1194, 2013. 2, 6, 7
  7. 7.C. Chen, S. McCloskey, and J. Yu. Image splicing detection via camera response function analysis. In Proc. IEEE Conf. Comput. Vis. Pattern Recogn., pages 5087–5096, 2017. 2
  8. 8.J. Chen, X. Kang, Y. Liu, and Z. J. Wang. Median filtering forensics based on convolutional neural networks. IEEE Signal. Proc. Let., 22(11):1849–1853, 2015. 1, 2
  9. 9.J. Chen, X. Liao, W. Wang, and Z. Qin. A features decoupling method for multiple manipulations identification in image operation chains. In Proc. IEEE Int. Conf. Acoust. Speech Signal Process, pages 2505–2509, 2021. 3
  10. 10.Y. Chen, J. Li, H. Xiao, X. Jin, S. Yan, and J. Feng. Dual path networks. In Proc. Neural Info. Process. Syst., pages 4470–4478, 2017. 8
  11. 11.D. Cozzolino, G. Poggi, and L. Verdoliva. Splicebuster: A new blind image splicing detector. In Proc. IEEE Int. Workshop Inf. Forensics Secur., pages 1–6, 2015. 1, 2
  12. 12.D. Cozzolino and L. Verdoliva. Noiseprint: a cnn-based camera model fingerprint. IEEE Trans. Inf. Forensics and Security, 15(1):114–159, 2020. 1, 2, 6, 7, 8
  13. 13.L. Van der Maaten and G. Hinton. Visualizing data using t-sne. Journal of Mach. Learn. Res, 9(11), 2008. 5
  14. 14.J. Dong, W. Wang, and T. Tan. Casia image tampering detection evaluation database. In IEEE China Summit Inter. Conf. Signal Info. Proc., pages 422–426. IEEE, 2013. 2, 6, 7
  15. 15.T. Gloe and R. Bohme. The dresden image database for benchmarking digital image forensics. J. of Digit. Forensic Pract., 3(2-4):150–159, 2010. 6
  16. 16.K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proc. IEEE Conf. Comput. Vis. Pattern Recogn., pages 770–778, 2016. 4
  17. 17.Y. Hsu and S. Chang. Detecting image splicing using geometry invariants and camera characteristics consistency. In IEEE Inter. Conf. Multim. Expo, pages 549–552. IEEE, 2006. 2, 6, 7
  18. 18.M. Huh, A. Liu, A. Owens, and A. A. Efros. Fighting fake news: image splice detection via learned self-consistency. In Proc. Eur. Conf. Comput. Vis., pages 101–117, 2018. 1, 2
  19. 19.A. Islam, C. Long, A. Basharat, and A. Hoogs. Doa-gan: Dual-order attentive generative adversarial network for image copy-move forgery detection and localization. In Proc. IEEE Conf. Comput. Vis. Pattern Recogn., pages 4676–4685, 2020. 2
  20. 20.X. Kang, M. C. Stamm, A. Peng, and K. R. Liu. Robust median filtering forensics using an autoregressive model. IEEE Trans. Inf. Forensics and Security, 8(9):1456–1468, 2013. 1, 2
  21. 21.A. Li, Q. Ke, X. Ma, H. Weng, Z. Zong, F. Xue, and R. Zhang. Noise doesn’t lie: towards universal detection of deep inpainting. In Proc. Int. Jt. Conf. AI, pages 1–7, 2021. 1, 2
  22. 22.H. Li, W. Luo, and J. Huang. Localization of diffusion-based inpainting in digital images. IEEE Trans. Inf. Forensics and Security, 12(12):3050–3064, 2017. 1, 2
  23. 23.Y. Li and J. Zhou. Fast and effective image copy-move forgery detection via hierarchical feature point matching. IEEE Trans. Inf. Forensics and Security, 14(5):1307–1322, 2019. 1, 2
  24. 24.X. Liao, K. Li, X. Zhu, and K. J. Liu. Robust detection of image operator chain with two-stream convolutional neural network. IEEE J. Sel. Top. Signal Process., 14(5):955–968, 2020. 3
  25. 25.T. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollar, and C. L. Zitnick. Microsoft coco: common objects in context. In Proc. Eur. Conf. Comput. Vis., pages 740–755, 2014. 6
  26. 26.S. Lyu, X. Pan, and X. Zhang. Exposing region splicing forgeries with blind local noise estimation. Int. J. Comput. Vis., 110(2):202–221, 2014. 1, 2
  27. 27.O. Mayer and M. C. Stamm. Forensic similarity for digital images. IEEE Trans. Inf. Forensics and Security, 15(1):1331–1346, 2020. 1, 2, 6, 7, 8
  28. 28.O. Ronneberger, P. Fischer, and T. Brox. U-net: convolutional networks for biomedical image segmentation. In Proc. Int. Conf. Med. Image Comput. Computer-Assisted Int., pages 234–241. Springer, 2015. 4, 6
  29. 29.A. G. Roy, N. Navab, and C. Wachinger. Recalibrating fully convolutional networks with spatial and channel “squeeze and excitation” blocks. IEEE Trans. Medical Imaging, 38(2):540–549, 2018. 6
  30. 30.S. Ruder. An overview of gradient descent optimization algorithms. arXiv preprint arXiv:1609.04747, 2016. 5
  31. 31.R. Shin and D. Song. Jpeg-resistant adversarial images. In Proc. Neural Info. Process. Syst. Workshop, pages 1–6, 2017. 4
  32. 32.W. Sun, J. Zhou, L. Dong, J. Tian, and J. Liu. Optimal pre-filtering for improving facebook shared images. IEEE Trans. Image Process., 30(1):6292–6306, 2021. 2
  33. 33.W. Sun, J. Zhou, Y. Li, M. Cheung, and J. She. Robust high-capacity watermarking over online social network shared images. IEEE Trans. Circuits Syst. Video Technol., 31(3):1208–1221, 2020. 2, 3, 6
  34. 34.W. Sun, J. Zhou, R. Lyu, and S. Zhu. Processing-aware privacy-preserving photo sharing over online social networks. In ACM Int. Conf. Multimedia, pages 581–585. ACM, 2016. 2
  35. 35.C. Szegedy, W. Zaremba, I. Sutskever, J. Bruna, D. Erhan, I. Goodfellow, and B. Fergus. Intriguing properties of neural networks. In Proc. Int. Conf. Learn. Representat., pages 1–10, 2014. 2, 5
  36. 36.H. Wu and J. Zhou. Iid-net: image inpainting detection network via neural architecture search and attention. IEEE Trans. Circuits Syst. Video Technol., pages 1–14, 2021. 1, 2
  37. 37.Y. Wu, W. AbdAlmageed, and P. Natarajan. Mantra-net: manipulation tracing network for detection and localization of image forgeries with anomalous features. In Proc. IEEE Conf. Comput. Vis. Pattern Recogn., pages 9543–9552, 2019. 1, 2, 6, 7, 8
  38. 38.J. You, Y. Li, J. Zhou, Z. Hua, W. Sun, and X. Li. A transformer based approach for image manipulation chain detection. In ACM Int. Conf. Multimedia, pages 3510–3517. ACM, 2021. 3
  39. 39.J. Zhong and C. Pun. An end-to-end dense-inceptionnet for image copy-move forgery detection. IEEE Trans. Inf. Forensics and Security, 15(1):2134–2146, 2020. 1, 2
  40. 40.P. Zhou, X. Han, V. Morariu, and L. Davis. Two-stream neural networks for tampered face detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recogn. Workshop, 2017. 2
  41. 41.P. Zhou, X. Han, V. I. Morariu, , and L. S. Davis. Learning rich features for image manipulation detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recogn., pages 1907–1915, 2018. 2
  42. 42.P. Zhuang, H. Li, S. Tan, B. Li, and J. Huang. Image tampering localization using a dense fully convolutional network. IEEE Trans. Inf. Forensics and Security, 16(1):2986–2999, 2021. 1, 2, 6, 7, 8

Citation

MLA
Wu, H., et al. “Robust Image Forgery Detection over Online Social Network Shared Images”. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 13430–39, https://doi.org/10.1109/CVPR52688.2022.01308.
APA
Wu, H., Zhou, J., Tian, J., & Liu, J. (2022). Robust Image Forgery Detection over Online Social Network Shared Images. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 13430–13439. https://doi.org/10.1109/CVPR52688.2022.01308
Chicago
Wu, H., J. Zhou, J. Tian, and J. Liu. 2022. “Robust Image Forgery Detection over Online Social Network Shared Images”. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 13430–39. https://doi.org/10.1109/CVPR52688.2022.01308.
Harvard
Wu, H. et al. (2022) “Robust Image Forgery Detection over Online Social Network Shared Images”, 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, pp. 13430–13439. Available at: https://doi.org/10.1109/CVPR52688.2022.01308.
Vancouver
1. Wu H, Zhou J, Tian J, Liu J (2022) Robust Image Forgery Detection over Online Social Network Shared Images. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, pp 13430–13439

BibTeX

@inproceedings{Wu_2022, title={Robust Image Forgery Detection over Online Social Network Shared Images}, url={http://dx.doi.org/10.1109/CVPR52688.2022.01308}, DOI={10.1109/cvpr52688.2022.01308}, booktitle={2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, publisher={IEEE}, author={Wu, Haiwei and Zhou, Jiantao and Tian, Jinyu and Liu, Jun}, year={2022}, month=June, pages={13430–13439} }
Metadata:Crossref

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF
License: IEEE