TensorFlow and Go
This year I helped organize several online security challenges, one of which is Blacklight. Among the things I was asked to do, was creating a POC for a specific challenge, to prove that it’s possible to solve in a reasonable time. That challenge was one I face occasionally in my everyday life, not always with success: break a captcha.
The task that requires breaking the captcha is disabling a security camera, to break into a room, without the security camera capturing your face. Here is how it looked before:
The provided information was the saved model used for captcha recognition in the binary ProtoBuf format, and a link to the camera control panel.
An input of a TensorFlow model requires doing some TensorFlow!
A Few words about TensorFlow
TensorFlow is an opensource software for Machine Intelligence, used mainly for machine learning applications such as neural networks.
TensorFlow runs computations involving tensors, and there are many sources to understand what a Tensor is. This article is definitely not a sufficient one, and it only holds the bare minimum to make sense of what the code does. Tensors are awesome and complex mathematical objects, and I encourage you to take the time to learn more about them.
For our purposes, here is the explanation from the TensorFlow website:
1

A tensor is a generalization of vectors and matrices to potentially higher dimensions. Internally, TensorFlow represents tensors as ndimensional arrays of base datatypes. 
A tensor is defined by the data type of the value(s) it holds, and its shape, which is the number of dimensions, and number of values per dimension.
The flow
part in TensorFlow comes to describe that essentially the graph (model) is a set of nodes (operations), and the data (tensors) “flow” through those nodes, undergoing mathematical manipulation. You can look at, and evaluate, any node of the graph.
A Few words about TensorFlow+Go
On the official TensorFlow website, you can find a page dedicated to Go, where it says “TensorFlow provides APIs that are particularly wellsuited to loading models created in Python and executing them within a Go application.” It also warns that the TensorFlow Go API is not covered by the TensorFlow API stability guarantees. To the date of this post, it is still working as expected.
When going to the package page, there are 2 warnings:
1) The API defined in this package is not stable and can change without notice.
2) The package path is awkward: github.com/tensorflow/tensorflow/tensorflow/go
.
In theory, the Go APIs for TensorFlow are powerful enough to do anything you can do from the Python APIs, including training. Here is an example of training a model in Go using a graph written in Python. In practice, some of tasks, particularly those for model construction are very low level and certainly not as convenient as doing them in Python. For now, it generally makes sense to define the model in TensorFlow for Python, export it, and then use the Go APIs for inference or training that model.[1] So while Go might not be your first choice for working with TensorFlow, they do play nice together when using existing models.
Let’s break into this page
The parts of the page I was facing seemed pretty familiar to your regular captchaprotected form:
PIN Code
 bruteforceCaptcha
 use the model
So my TO DOs were:
 1. Build a captcha reader
 2. While not logged in:
 2.1 Generate the next PIN code
 2.2 Get captcha text for current captcha image
 2.3 Try to log in
SavedModel CLI
From the website: SavedModel is the universal serialization format for TensorFlow models.
So our first step would be figuring out the input and output nodes of the prediction workflow. SavedModel CLI is an inspector for doing this. Here’s the command and its output:
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 
$ saved_model_cli show dir <PATH> all MetaGraphDef with tagset: 'serve' contains the following SignatureDefs: signature_def['serving_default']: The given SavedModel SignatureDef contains the following input(s): inputs['input'] tensor_info: dtype: DT_STRING shape: unknown_rank name: CAPTCHA/input_image_as_bytes:0 The given SavedModel SignatureDef contains the following output(s): outputs['output'] tensor_info: dtype: DT_STRING shape: unknown_rank name: CAPTCHA/prediction:0 Method name is: tensorflow/serving/predict 
What we learn from this are the node names.
Input node: CAPTCHA/input_image_as_bytes
,
Output node: CAPTCHA/prediction
.
Captcha
Now let’s load the model, using func LoadSavedModel(exportDir string, tags []string, options *SessionOptions) (*SavedModel, error)
. The function takes 3 arguments: path, tags and seesion options. Explaining tags and options can easily take the entire post and will shift the focus, so for our purpose I used the convention {"serve"}
, and provided no session options.


Then get the captcha from the web page, and run it through the model. First, define the output of an operation in the graph (model+node) and its index.


Create a new tensor. The input can be a scalar, slices, or array. As we want to predict a captcha, we’ll need 1 dimension with 1 element, of type string.


Set a map from the operation we will apply to the input it will be applied on.


Get the output from the prediction operation into this output struct.


Run the data through the graph and receive the output  the captcha prediction.


Here is how this looks like:
Generate a PIN code
The PIN code is made of 4 digits, so we’ll go over all the combinations. Additionally, in each iteration the saved model is required for the prediction operation, and of course some logs.


Try to log in
Once all values are there  the current value of the PIN code in the loop and the captcha prediction  let’s POST that request to the login page.


If the captcha prediction failed, run the prediction again, and retry with the same PIN code.


The parseResponse
function checks and reports whether the website response is a success or one of the failure messages, which I found by manually trying combinations of guessing a PIN code and correct and wrong captcha translations.


The rest of the code
To complete this code, let’s add everyones favorites: cookies and logging. Generating the captcha starts a new session, and in order to use the predicted captcha in the same session, we will open a cookie jar. Even though it’s the first time I am writing about cookies publicly, I will spare cookie jokes, as part of the Christmas spirit.


And here is how it looks when it’s all composed together.
To wrap this up
TensorFlow has many great models which can be used with Go. Here is a great list of those.
Online challenges can be an awesome way to learn, whether it’s coding, security or sports. The combination of putting in practice your knowledge and having a mission creates a fun environment where you can work on improving your skills. Consider joining such a challenge as your new year’s resolution.
Thanks a lot to Ed for reviewing this PR. Also thanks to Asim Ahankar from the TensorFlow team for pointing out it is possible to train models with Go, as updated in [1]. We will collaborate further to make the documentation around this more accessible.
If you want to chat more about this, tweet me, or meet me at Gophercon Iceland!