Improving Robustness of Image Tampering Detection for Compression

CORESA 2018

Transcript

1. Improving Robustness of Image Tampering Detection for Compression Boubacar Diallo, Thierry Urruty, Pascal Bourdon and Christine Fernandez-Maloigne XLIM Research Institute (UMR CNRS 7252), University of Poitiers, France CORESA 2018 - Poitiers, France November 14th, 2018
2. Introduction
3. Image forensic: global context • Tampered image generation processes widely available. • Most common manipulations: Copy-move, Splicing and Removal. • Distribute misleading messages within the society: e.g: Mass manipulation, Cybercriminality, Tampering or removing of judicial evidence... 1
4. Image forensic: global context • Camera Model Identification (CMI). • Which camera brand took this picture? What model? Specific device? • Crucial for criminal investigations and legal proceedings. • Using the convolutional neural networks (CNN) to learn camera identification features. 2
5. Our contribution • Many authenticity constraints on tampering detection algorithms. • With unpredictable changes caused by manipulations: e.g compression, noising, resizing and/or filtering. • A robust framework which contributes to improving camera model identification and image tampering detection for such manipulations. • Robust framework against compression. 3
6. Proposed Method
7. Global framework 1. A Deep Learning approach to identify camera models with different transformations. 2. A robust solution for tampering image detection heeding these transformations. 4
8. Camera Model Identification (CMI) • All original images have to be duplicated with transformed versions of itself. • CNN contains 11 layers: 4 Conv, 3 Maxpooling, 2 fully-connected (ReLU and Softmax). 5
9. Tampering image detection • A clustering algorithm estimates a tampering mask M. • The final output M is a binary mask: • Black parts indicate patches belonging to the pristine region. • White parts indicate the forged patches. 6
10. Experiments and results
11. Datasets and evaluation criteria • To evaluate Camera model identification: Dresden dataset It contains more than 13, 000 images of 18 different camera models. • To evaluate tampering detection algorithm: Two sets of altered data: Represent a set of ”known” data from test dataset and an ”unknown” dataset from ”unknown” cameras. Each contains 500 pristine images and 500 tampered images generated following in 1. • Evaluation criteria: Accuracy and True positive rate (TPR). 1Bondi, Luca, et al. ”Tampering Detection and Localization Through Clustering of Camera-Based CNN Features.” CVPR Workshops. 2017. 7
12. Training with compressed dataset • Quality of the input data should respect the desired application. • ”Original” sets are compressed with quality factors of 90%, 80%, and 70%. • Trained CNN: CNN90, CNN80, CNN70 and CNNm respectively for 90%, 80%, 70% and mixed compressed data. • This step is of great importance to obtain good performance (Experiments). 8
13. Influence of compression on CMI The performance of Bondi et al. approach decreases dramatically with compression quality factors inferior (QF = 90, 80, 70 etc.) The CNN trained only on ”Original” image for CMI is not robust to compression. [Bondi et al. CVPR 2017] Accuracy comparison curves of CMI 9
14. Influence of compression on CMI Compression of datasets with quality factor values ranging from 70 to 100 and training of them. Data Augmentation: by the horizontal and vertical flipping of training dataset. An average accuracy for CMI. Accuracy comparison curves of CMI 10
15. Influence of compression on CMI Higher accuracy for CNN trained model on a specific compression QF. Our framework gives performing results on average of all mixed compressed test images. Accuracy comparison curves of CMI 11
16. Influence of compression on Tampering image detection Our performance outperforms for any other compression quality factor. Dataset Compression Accuracy TPR Bondi 2 CNNm Bondi CNNm Known Original 0.84 0.77 0.9 0.83 90% 0.56 0.72 0.47 0.68 80% 0.52 0.65 0.24 0.48 70% 0.52 0.61 0.15 0.38 Unknown Original 0.79 0.7 0.84 0.68 90% 0.56 0.63 0.38 0.46 80% 0.52 0.57 0.17 0.30 70% 0.51 0.56 0.11 0.26 2Bondi, Luca, et al. ”Tampering Detection and Localization Through Clustering of Camera-Based CNN Features.” CVPR Workshops. 2017. 12
17. Influence of compression on Tampering image detection An example of a tampered image with the detected area. 13
18. Conclusion and perspectives
19. Conclusion and perspectives • Study of the compression effect on camera model identification and tampering image detection • A very good match between the analytical and experimental results validates our approach. • Study especially important for forensic researchers because of the higher compression rates applied to images on Internet. • In the perspectives: Investigate the neural network activation kernels to better understand the artifacts that help to identify camera models. 14
20. Questions / Answers Thank you for your attention
21. Questions / Answers Thank you for your attention
22. Influence of compression on Tampering image detection ROC and AUC study the effect of compression quality factor on our framework • The loss of accuracy is closely linked to the quality of an image.