The general idea is that there are many noise sources which technology normally needs to route around. To find random number sources, you reverse this.
Aside of that, you get no direct hardware access to measure and even if you would, there is still the timer resolution problem.
Very recent browsers provide an API to get strong random numbers, but this is even less widely available than web sockets.
So either you take the mouse movement or you use something like Java or Flash, or you use bad numbers.
The JS time function returns epoch milliseconds, not seconds. At the scale of network latency (hundreds if not thousands of milliseconds), the lower order bits should be effectively uniformly distributed. Resolution isn't your problem, the number of usable bits per network call is.
x=new Date().getTime(); for(j=0;j<10000;j++); y=new Date().getTime(); for(j=0;j<10000;j++); z=new Date().getTime(); alert(x+' '+y+' '+z)
gives me high resolution timestamps:1311577733069 1311577733083 1311577733099
you could also point the camera at a flat white surface - that would give you a lot more noise (so much you probably don't need to worry about subtracting anything, if you just take the least significant bits from each pixel)
[i hope i'm right - i wrote astronomy image reducing software for a living]
Easiest way is probably with a webcam, which will also give you a pretty good bitrate. The general schema is to read out the bias noise. You do this by blocking out any incoming light to get dark noise + bias noise. Dark noise is an assumed static shift due to CCD characteristics, so read N frames of this and look for the median signal. Subtract that off and you've got the bias, or at least something that's close enough for a weekend project.
Better version, start reading up on the math of noise sources in whatever device you want to use for a sensor. Also do an analysis of N samples to see how close the result comes to the expectation value given the type of noise involved and what deviation is expected at N samples.