RED GREEN BLUE
11111111 00000000 00000000
A change in the most significant bit of any of these colours would be massive. 01111111 is half as bright as 11111111.But a change in the least significant bit(s) is barely noticeable. You probably wouldn't notice the difference between 11111111 and 11111110, especially as you don't experience individual colour channels, you see a blended colour. And when you're looking at something with thousands of pixels you simply won't spot it.
So, stick your secret message in the low bits:
RED GREEN BLUE
11111XXX 00000XXX 00000XXX
You now get 9 bits of data per pixel. Or less (for a less noisy image) or more (for a more noisy image). For a 640x480 image, you now have 345.6KB of data. And if you use encrypted data, it won't even seem like data if someone looks (plausible deniability). Just noise.Techniques like this were used on 4chan to hide child porn in high-resolution photographs. This the "mods are asleep, post hi-res" meme.
Turns out that the least significant bits of a pixel are not uniformly distributed. The obvious example of failure is that you expect more pixels with value 255 than 254, since any overexposure will be stuck at 255. The rest is obvious when you look at a histogram of pixel values; as the number of LSBs used increases you see the curve become stepped.
LSB replacement steganography is easily detectable at about a ratio of about 0.01 bits per pixel. The smarter plan is to add or subtract 1 randomly (or do nothing) in order to match the LSB to what you want; this symmetry is harder to detect (but still very possible, the methods are just more complicated).
The reason you encrypt the data is not just for data security, it is so that the embedder does not have to worry about the data that they are embedding screwing everything up.
tl;dr LSB steganography does not give you any plausible deniability because it's really easy to detect in any type of image at very low embedding levels due to inherent limitations of the process. If you want to use it, you'd be safe with a single message at perhaps a level of 0.005 bits per pixel, but then in a ten megapixel image you'd get 150kB of data which hardly seems worth it when you have to transmit losslessly. Even then, you'd still be screwed at a theoretical level with Ker's square root law.
Still, I think LSB replacement steganography is useful if only to provide an example of how it steganography can be done, even if it's not particularly sophisticated.