A NEW ADVERSARIAL EMBEDDING METHOD FOR ENHANCING IMAGE_REVIEW2.pptx

This a ppt on research paper

1. A NEW ADVERSARIAL EMBEDDING METHOD FOR ENHANCING IMAGE STEGANOGRAPHY TEAM MEMBERS : K.JAHNAVI(Y21IT051) B.DEVA SIVA NAGA SAI(Y21IT012) B.MANIKANTA(Y21IT010) Guide : Dr.A.Yaswanth kumar

2. CONTENTS : Problem statement Block diagram Proposed method Discussion Conclusion References

3. PROBLEM STATEMENT In recent years, image steganography has become an increasingly crucial aspect of information security, facilitating covert communication by embedding secret data within images imperceptibly. However, existing steganographic methods often face challenges in terms of robustness and detectability. An adversarial embedding method is proposed to aim at enhancing the effectiveness and stealthiness of image steganography.

4. BLOCK DIAGRAM

5. PROPOSED METHOD The proposed method aims to enhance the security performance of a given steganography F via adjusting some original embedding costs within covers. To this end, we firstly train a CNN-based steganalyzer NT based on covers and the corresponding stegos produced by F, and then generate new stegos based on adversarial examples to mislead NT . The framework of the proposed method includes two steps Steganalytic Network Pre- Trainin g Stego Generation

6. 1) STEGANALYTIC NETWORK PRE-TRAININ G We firstly collect a cover image set CT , and obtain the corresponding cost map set ρCT using a steganography F to be enhanced, and then generate the corresponding stego image set ST using STC at a given embedding payload. Finally, we can obtain a CNN-based steganalytic network NT by training the cover set and stego set, i.e., CT and ST . The proposed method is non-iterative since all gradients are calculated based on the fixed steganalyzer NT via back propagation.

7. 2) STEGO GENERATION In this step, we collect another cover image set CA, where CA ∩CT = ∅. For any cover image c ∈ CA, we firstly compute its original cost map ρ0 by the steganography F. Then we generate n random messages with a fixed payload, and embed them respectively into the cover c to generate n different stegos (i.e., s1,s2,..., sn ) using STC based on the same embedding costs ρ0. Then we feed cover c and its n different stegos into the pre-trained network NT to obtain a gradient map Gc for cover c and multiple gradient maps (i.e., Gs1 , Gs2 ,··· , Gsn ) for the n stegos using back propagation algorithm.

8. CONT… For a given input image x (cover c or stego s1,s2,..., sn ), the corresponding gradients Gx is calculated based on the formula After calculating gradients of a cover and stegos , we then select embedding units within cover for subsequent modifica tion .

9. Algorithm for Selecting Embedding Units to Be Modified INPUT: Gradient maps Gc & Gs1 ,..., Gsn ; original cost map ρ0 of a cover c; the parameter p1, p2 for thresholding gradients and embedding costs; the number n of stegos . OUTPUT: Coordinate sets of selected embedding units in positive and negative directions M+ and M−. PROCEDURE: M+ n = {( i , j) | Gc( i , j) > 0∩Gsk ( i , j) > 0, k = 1,..., n} M− n = {( i , j) | Gc( i , j) < 0∩Gsk ( i , j) < 0, k = 1,..., n} M(p1,p2) = Low ρ0 ( p1) ∩ Top|Gc |(p2) M+ = M+ n ∩ M(p1,p2) M− = M− n ∩ M(p1,p2) Return: M+,M−

10. CONT… For the selected units in positive and negative modification directions (i.e., M+ and M−), we update their original embedding costs ρ0 as follows: Based on the modified embedding costs ρ, we finally get the resulting stego image s for a cover c using STC .

11. RESULT Selected embedding units to be modified. The enhanced steganography is HILL with a payload 0.4 bpp , and p1 = 0.7, p2 = 0.1, n = 5.

12. DISCUSSION To achieve more convincing results, we randomly split CT and CA three times and report the average results in the following. Parameter Selectio n Security Evaluation on Pre-Trained Steganalyzer Security Evaluation on Re-Trained Steganalyzers Comparison With Related Works Analysis on Cost Modification Rate

13. CONT… A) PARAMETER SELECTIO N: There are three important parameters, that is, the number of stegos n and two percentages p1 and p2 for thresholding the costs and gradients. Parameter Selection for (p1, p2): Parameter Selection for n Parameter Selection for α

14. CONT… B) SECURITY EVALUATION ON PRE-TRAINED STEGANALYZER: The proposed method can effectively mislead the targeted steganalyzer in all cases. The average improvements would increase with increasing the embedding payloads. For instance, the average improvement is around 9% when the embedding payload is 0.1 bpp , while it becomes over 16% when the embedding payload is 0.4 bpp . C) SECURITY EVALUATION ON RE-TRAINED STEGANALYZERS: The proposed method can effectively mislead a pre-trained steganalyzer . When the proposed method is exposed, however, the detectors would re-train the targeted network (i.e., Deng-Net) based on the resulting stegos , and/or employ other steganalyzers for security evaluation.

15. CONT… D) COMPAISON WITH RELATED WORKS:

16. CONT…

17. CONT…

18. CONT… COMPARISON OF EXECUTION TIME: The execution time of ADV-EMB based method is relatively longer (over 3 times) than the proposed method and AEN, since it has to search a good partition parameter p by iteratively executing the main process of ADV-EMB. On average, the proposed method takes around 0.21 second to generate a stego , which is acceptable in practical application.

19. CONT… F) ANALYSIS ON COST MODIFICATION RATE: Unlike most related works that modify almost all or a random part of embedding units in an image, the proposed method tries to select and modify some embedding costs to enhance the security of existing steganography.

20. CONCLUSION Unlike existing related works, the proposed steganography can make full use of the gradients from cover and generated multiple stegos to select and adjust original embedding costs within covers. Besides gradient signs, the proposed method proposes to combine the amplitudes of gradients and original embedding costs of cover for selecting candidate embedding costs for subsequent modification . Furthermore , the security performances evaluated on different image database show that the generalization of the proposed method is good.

21. REFERENCES T. Song, M. Liu, W. Luo, and P. Zheng, “Enhancing image steganography via stego generation and selection,” in Proc. IEEE Int. Conf. Acoust ., Speech Signal Process. (ICASSP), Jun. 2021, pp. 2695–2699. D. Smilkov , N. Thorat , B. Kim, F. B. Viégas , and M. Wattenberg, “ SmoothGrad : Removing noise by adding noise,” CoRR , vol. abs/1706.03825, pp. 1–10, Jun. 2017. C. Xie et al., “Improving transferability of adversarial examples with input diversity,” in Proc. IEEE/CVF Conf. Comput . Vis. Pattern Recognit . (CVPR), Jun. 2019, pp. 2730–2739. B. Chen, W. Luo, P. Zheng, and J. Huang, “Universal stego postprocessing for enhancing image steganography,” J. Inf. Secur . Appl., vol. 55, Dec. 2020, Art. no. 102664. M. Yedroudj , M. Chaumont, F. Comby , A. O. Amara, and P. Bas, “Pixels-off: Data-augmentation complementary solution for deeplearning steganalysis,” in Proc. ACM Workshop Inf. Hiding Multimedia Secur ., Jun. 2020, pp. 39–48.