no algorithm can work about that but many statistical formula can do that

You don't define «key» og «key length».  If you mean extracting the key which is used to choose pixels/coefficients for embedding, that is only possible if the embedded message has a lot of structure (or you mount a known plaintext attack) and even then it is harder than breaking a long permutation cipher.  If the embedded message is compressed, encrypted or otherwise random-looking, it is nigh impossible (by the definition of random-looking).

Your question might make sense if you really mean «message length».  Then you could use any kind of regression algorithm (e.g. support-vector regression), in combination with some feature set for machine learning as used for yes/no steganalysis.  (You might want to check out my book on Machine Learning in Image Steganalysis.)

I won't split any hairs over the difference between an algorithm and a formula :-)

Thank you Dr. Hans Georg Schaathun,

Yes I mean Msg length and embedding key, as they are the factors that influence the success of embedding alg., since they are its input

In my question, I am looking for the weakness points in embedding algorithms of image steganography. One idea to cipher a msg then embed according to some key.

I would advice you to disregard the most blatant and trivially correctable weaknesses. Embedding an unencrypted message is just silly. Encryption is easy and computationally less costly than embedding, so there is absolutely know reason why a steganographer should embed unencrypted text.

Estimating the message length is possible and potentially useful under certain assumptions. It can work even with encrypted messages. Some work has been done in the past, but not so much recently I think.

The embedding key sounds like a lost cause.  I don't see why you want it, and with a reasonable steganographer I cannot see that it is possible,

Thank you Hans Georg Schaathun , the embedding key leads to discover the bits that are changed in the image