Basically it's analysis of a cryptographic cypher by the means of finding a relationship between the difference in the input data and the output data. Ideally, the slightest difference in input data (cleartext), even a single bit, should produce a completely different cypthertext. However, if the cypher is not well-designed, a correlation between the two resulting cyphertext might be observed. This correlation in turn might be exploited to find out the key. This, obviously requires a chosen-cleartext-attack; which means the attacker should have access to the encryption mechanism, and thus use it to encrypt any number of cleartexts he or she choses. Therefore, this kind of attack is generally applied against temper-proof hardware which one could use but not get hands on the key itself.
As per AES, see this excerpt from the Wikipedia article on differential cryptanalysis:
For example, if a differential of 1 => 1 (implying a difference in the LSB of the input leads to a output difference in the LSB) occurs with probability of 4/256 (possible with the non-linear function in the AES cipher for instance) then for only 4 values (or 2 pairs) of inputs is that differential possible. Suppose we have a non-linear function where the key is XOR'ed before evaluation and the values that allow the differential are {2,3} and {4,5}. If the attacker sends in the values of {6, 7} and observes the correct output difference it means the key is either 6 xor K = 2 or 6 xor K = 4, meaning the key is either K = {2,4}. In essence, for an n-bit non-linear function one would ideally seek as close to 2-(n-1) as possible to achieve differential uniformity. When this happens, the differential attack requires as much work to determine the key as simply brute forcing the key.
Here are a few links for warming up:
- A simplified AES algorithm and its linear and differential cryptanalyses
- Impossible differential cryptanalysis of Mini-AES
- Low Complexity Differential Cryptanalysis and Fault Analysis of AES
And here are a couple of better articles on the subject:
- A Tutorial on Linear and Differential Cryptanalysis
- A stochastic model for differential side channel cryptanalysis
Not top quality links, and especially not as practice-oriented as you seem to be looking for, but I think these last two could give you a bit of grip on the subject.