Information Hiding - Science topic

The proposed algorithm has binary codes and pixels inside an image

That is what I expected: 8 bit gray scale makes max. 48 dB from the pixel values (20 * log (MaxI). (Less, if you have images not reaching the 0xFF white level.)

Anything beyond that comes from very small MSE-Values. MSE 0.01 would add 11.6 dB, but I guess you have much lower values in some images. Could be from a lack of edges at all, could be from "soft edges".

So, everything is according to the theory. When looking at an explanation, the term spitting the PSNR into terms for image content and MSE contribution should help assessing the "meaning" of PSNR values below/above 100 dBs.

Good luck and have fun

you can search on web with "reversible data hiding" because it is the same word.

It seems that many of the responders use `steganography' to refer to what I would call robust watermarking.  I will consider steganography where the attacker's objective is to detect the presence of a secret communication.  Ingemar Cox has discussed the distinction in illuminating detail.

You can only study or describe the robustness relative to a certain context, i.e. assuming a specific application scenario and a specific attack scenario.

Aspects of the context includes

1.  Choice of cover data.

2.  Size of the cover data and size of the payload.  Note that the problem does not scale linearily, as you will see from Andrew Ker's work on the square root law of steganography.

3.  Steganalysis algorithm used by the attacker.  

4.  Frequency of steganogrammes in the medium where you distribute your steganogrammes, as this effects the likelihood that a given message is a steganogramme both prior to and after steganalysis.

Virtually all work on steganalysis and steganographic robostness has focused on a completely artificial context (lab conditions) with little regard for real application. Typically that means a given pool of covers, a small set of attacks, a given relative payload, and assuming that only a single steganogramme is ever intercepted. Andrew Ker's work on batch steganography and pooled steganalysis is a notable exception.  I also discuss the challenge of addressing real life problems based on lab experiments in my book on Machine Learning in Image Steganalsysis, but I do not have a solution.

Image quality assessment metrics suggested by previous contributors are MSE and PSNR. These metrics are simple to calculate,mathematically convenient, and have clear physical meanings.  However these IQ assessments lack two critical features: 1) Ability to image similarity across stego image of varying distortion types; 2.) They are not very well matched to perceived visual quality. A new approach to assess image quality independent of the type and size of pixel size of concerned stego images is the estimation of Structural Similarity indices using  Structural Similarity Index metric(SSIM)

The author might need to see our investigation of SSIM computation of dissimilar stego images in secure e-voting application in this link:

(Combining methods are always being asked in conjunction with crypto, typically: encryption, compression, error correction, etc.)

You can combine in many ways, take Gunter's two ways and you can also iterate.

However, the more you combine the more you pay in performance, of course! (and this was noted above). So combination is often not practical.

Also, if your cipher is good why embed the message in a file first and then encrypt it rather then encrypting the message directly? It is a technique but does it make sense (ciphertext entropy will always reveal it is an encryption, and will defeat the basic trait of stego.)?

The main advantage of stego is its subliminal nature (hiding without being noticed), thus "encryption then stego" has some meaning (if your stego is broken [say by file dif] your message is still protected), while encrypting after stego has less use.

There are three methods of data hiding. In cryptography, we can openly announce that we added some data and the interceptor don't have key to open. In watermarking also, we can announce the same, but interceptor can attack te file to remove the data. We need to embed small data for identification. However, in steganography, we cannot announce to anybody, because if anyone is able to find the presence of data in image, they can remove the file. So our aim will not be achieved. So, in steganography, We must be able to exploit HVS (Human Visual System). This we can achieve using frequency domain because our eyes cannot detect high frequencies. So, either you can go for DCT or DWT. if you are using DCT, don't worry about lossy compression. You just have to aim to embed the data in low frequencies which you can decide that how much compression your file can handle. Check out papers in my profile on watermarking. You can check the tables mentioning the compression attacks.