Generative steganography framework. 

Contexts in source publication

Context 1

... generative steganography framework proposed in this paper is illustrated in Fig. 1. In this scenario, the sender directly creates a stego carrier using a generator with a message and a key. Here, the embedding algorithm becomes a stego sampling (generation) process. The secret key shared by both parties ensures the security of message extraction, and the degree of naturalness of the stego image determines the ...

Context 2

... to the nonconvexity of the models in the training scheme, we cannot guarantee that the generator will converge to a model that can enable perfect recovery the secret message from the steganographic image . Fig. 11(a) and (b) shows the relationships between the message extraction accuracy rate and the number of iterations with LWF. increases. Due to the randomness of pixel generation and the fact that the message constraint cannot be completely satisfied. As can be seen from the Fig.11(a) and (b), as the amount of embedding increases, the ...

Context 3

... to the randomness of pixel generation and the fact that the message constraint cannot be completely satisfied. As can be seen from the Fig.11(a) and (b), as the amount of embedding increases, the accuracy of message extraction decreases. ...

Context 4

... between the message extraction accuracy rate and the number of iterations with LWF. increases. Due to the randomness of pixel generation and the fact that the message constraint cannot be completely satisfied. As can be seen from the Fig.11(a) and (b), as the amount of embedding increases, the accuracy of message extraction decreases. As shown in Fig. 11(c), we also performed message embedding and extraction average accuracy at different BPIs for 1,000 images that were not in the training set. All the stego images were sampled over 3,000 iterations, starting with corrupted images with missing 99% regions of the image. As expected, the accuracy of message extraction increased as the BPI ...

Context 5

... at different BPIs for 1,000 images that were not in the training set. All the stego images were sampled over 3,000 iterations, starting with corrupted images with missing 99% regions of the image. As expected, the accuracy of message extraction increased as the BPI value increased. The receiver was able to recover more than 70% at BPI = 7. Fig. 11(c) shows the relationship between the average error rate and the BPI. Compared with our previous work reported in [21], the message extraction stability is ...

Context 6

... capacity of our method is 1.49e-3~1.10e-2 bytes per pixel, where the lower and upper ends of this range correspond to image corrupted rates of 0.99 to 0.91, respectively, as shown in the last row of Table 1, In theory, the relative capacity of the proposed method can be higher, but under the message loss constraint in the message, as shown in Fig. 11, extraction accuracy will be very low, so the high embedding amount does not have practical value, the actual capacity should be calculated ...

Context 7

... traditional steganalyzers for the cover modification method, all 1,000 stego images and 1,000 normal images were generated over 1,000 iterations from corrupted images via the image inpainting process. The database was randomly divided into two halves: one for training and the other for testing. We averaged the performances over ten random splits. Fig. 12 is a plot showing the evolution of the testing error PE as a function of the payload from 0.01 bits per pixel (bpp) to 0.05 bpp when BPI = 7. For comparision, this plot also includes the corresponding results for HUGO [28] and HILL [29], which are considered to be advanced steganographic methods that minimize distortion by means of ...

Context 8

... results in Fig.12 show that steganography based on sampling can resist statistical steganographic analysis, mainly because the completed stego and normal images can be regarded as samples from the distribution pg. ...

Context 9

... results in Fig.12 show that steganography based on sampling can resist statistical steganographic analysis, mainly because the completed stego and normal images can be regarded as samples from the distribution pg. There is no pairwise relationship between the features extracted from the normal cover image and the stego image. As shown in Fig. 12, our method performs competitively with HUGO [28] and HILL [29] for cases with a low embedding rate. ...

Context 10

... Generative Steganography framework of this paper is shown in Fig.1 as follows: In this scenario, the sender create a stego carrier from a generator with message directly. The embedding algorithm actually turns into a stego sampling (generation) process. The secret key shared by both parties ensures the security of the message, the natural real degree of stego determines the security of the communication ...

Context 11

... from the traditional steganalyzer for cover modification method, all 1000 stego images and performance is averaged over ten random splits. In Figure 11, we plot the progress of the testing error PE as a function of the payloads from 0.1bpp to 0.5bpp (bits per pixel) with BPI = ...

Context 12

... to the non-convexity of the models in the training scheme, we cannot guarantee that the generator will converge to the model that can successfully recover the secret message from the steganographic image perfectly. Fig. 10a shows the relationship between the error rate of the message extraction and the number of iterations with different BPI (1-8). As shown in Fig.8. We do the message embedding and extraction at different BPI for 1000 images which not belonging to the training set. All stego images are sampled at 3000 iterations from corrupted images with 90% region missing. As expected, the accuracy of message extraction increased with the increase of BPI. The receiver was able to recover more than 95 % of messages sent by sender when BPI >3. Our scheme can perfectly decode the secret encrypted message from the steganographic image at BPI =8. In the Fig 10b, the relationship between the average error rate and BPI is given, compare with our work in [21], the stability of the message extraction is greatly improved. We steganalyze our digital Cardan grille method using blind steganalyzer for spatial domain and the ensemble classifier. 686-dimensional SPAM features [25] and 5404-dimensionnal SCRMQ1 features [26] with ensemble classifiers [27] implemented as random forests are used for this ...

Context 13

... to the non-convexity of the models in the training scheme, we cannot guarantee that the generator will converge to the model that can successfully recover the secret message from the steganographic image perfectly. Fig. 10a shows the relationship between the error rate of the message extraction and the number of iterations with different BPI (1-8). As shown in Fig.8. We do the message embedding and extraction at different BPI for 1000 images which not belonging to the training set. All stego images are sampled at 3000 iterations from corrupted images with 90% region missing. As expected, the accuracy of message extraction increased with the increase of BPI. The receiver was able to recover more than 95 % of messages sent by sender when BPI >3. Our scheme can perfectly decode the secret encrypted message from the steganographic image at BPI =8. In the Fig 10b, the relationship between the average error rate and BPI is given, compare with our work in [21], the stability of the message extraction is greatly improved. We steganalyze our digital Cardan grille method using blind steganalyzer for spatial domain and the ensemble classifier. 686-dimensional SPAM features [25] and 5404-dimensionnal SCRMQ1 features [26] with ensemble classifiers [27] implemented as random forests are used for this ...

Context 14

... SPAM feature does not work. This is mainly because SCRMQ1 is designed for color images, and SPAM is designed for grayscale images. We also give the error rate for different iterations. This is shown in the Figure 13 below. After dozens of iterations, SPAM and SCRMQ1 features maintain consistent performance. Experiments show that the resistance to statistical analysis, does not mean that the image generation quality is good enough, in fact, with the increase of the number of iterations, image visual distortion to reduce gradually, as shown in Figure 9. ...

Context 15

... SPAM (b) SCRMQ1 From the above experiments, it can be seen that the steganography based on sampling can resist the statistical analysis of the steganography, this is mainly due to the fact that, completed stego and normal images can be regarded as samples from the same distribution p g . The normal cover and stego does not have a pairwise relationship between the extracted features. As can be seen from the figure, our method has competitive performance with HUGO [28] and HILL [29] in the case of low embedding rate. Figure 12 shows the average classification error PE achieved with five different BPI at 0.1bpp with 1000 iterations. As the bit plane index increases, the security of our method decreases on ...