Changes

Additionally, this tutorial uses some terminology from previous tutorials, such as ''Hamming weight'' and ''substitution box''. If you don't know what these are, [[Tutorial B6 Breaking AES (Manual CPA Attack)]] might be an easier starting point.

All of the work in this tutorial will be done using Python scripts. All you need for this tutorial is a text editor and a command terminal that can run Python. In my example, I am editing a file called <code>manualTemplate.py</code> in a text editor and using the command

to execute the script. You can work with a full IDE if you prefer.

As in the previous tutorial, this tutorial requires two sets of traces. The first set is a large number of traces (1000+) with random keys and plaintexts, assumed to come from your personal copy of the device. The second is a smaller number of traces (~50) with a fixed key and random plaintexts, assumed to come from the sensitive device that we're attacking. The goal of this tutorial is to recover the fixed key from the smaller set of traces.

Note that this tutorial will explain how to attack a single byte of the secret AES key. It would be easy to extend this to the full key by running the code 16 times. The smaller attack is used to make some of the code easier to grasp and debug.

This section describes how to generate a template from the random-key traces. Our template will attempt to recognize the Hamming weight of the AES substitution box output. This choice of attack point limits our template - we will not be able to find the secret key in one attack trace - but it allows us to use a smaller amount of preprocessing. A more robust template attack (for instance, creating a template for each of the 256 possible key bytes) would require at least 10 times more data.

The traces recorded from the ChipWhisperer Capture program are saved as NumPy arrays. These files can be loaded using the <code>np.load()</code> function. We're interested in the traces, plaintext, and random keys used in our template captures, so we can load this data with the code:

Get in the habit of checking your data with some basic print statements or plots - it's a good sanity check to make sure that your arrays are the size they should be! If everything looks okay, comment out these checks and move on.

With the data in-hand, our next task is to group the data according to our model. We're attacking an intermediate result in the AES algorithm where

</pre>

After sorting the traces by their Hamming weights, we need to find an "average trace" for each weight. We can make an array to hold 9 of these averages:

</pre>

With 5 (or <code>numPOIs</code>) POIs picked out, we can build our multivariate distributions at each point for each Hamming weight. We need to write down two matrices for each weight:

* A mean matrix (<code>1 x numPOIs</code>) which stores the mean at each POI

Note that you may not have enough data to complete the covariance matrix - check out the Gotchas at the end of this tutorial if this step blew up on you.

Our template is ready, so we can use it to perform an attack now. We'll load our fixed-key traces and apply the template PDF to see how good each guess is, keeping a running total to check which subkey guesses are the best matches.

If you followed the instructions from the previous tutorial, you'll have a few dozen traces that use random plaintexts and a fixed key. We can load these in exactly the same way as the template traces:

<pre>

We're only attacking the first subkey, so let's see if we can get <code>43</code> to come up as the best guess.

The very last step is to apply our template to these traces. We want to keep a running total of <math>\log P_k = \sum_j \log p_{k,j}</math>, so we'll make space for our 256 guesses:

<pre>

so we successfully attacked the first subkey in 3 traces! This is a bit lucky - most of the subkeys tend to take a bit more data. Don't be surprised if your attack takes closer to a dozen trials.

One problem with template attacks is that they require a large amount of data to make good templates. Remember that every output of the AES substitution box is unique, so there is only one output with a Hamming weight of 0 and only one with a weight of 8. This means that, when using random inputs, there is only a 1/256 chance of a trace using a Hamming weight of 0 or 8.

If you ran into numerical problems while working through this tutorial, try recording another bigger data set. Instead of capturing 1000 template traces, try 5000 (on your coffee break), 10000 (on your lunch break), or 100000 (overnight). You'll probably find that the extra data makes the statistics work out better.

If you got lost, here's our full implementation:

<pre>