misterpotpot / shacrush

SHA-256 First Character Predictor

Objective

The objective of this project is to investigate whether a neural networks based model can reverse engineer the SHA-256 hashing algorithm. Initially, this neural network is designed to predict the first character of an 8 to 24 characters ASCII string based on its SHA-256 hash. This serves as a Proof of Concept, which we could expand upon if it looks promising. The ultimate goal is to reverse engineer complete ASCII strings of known lengths from their SHA-256 hashes, employing individual neural networks for each character. Subsequently, we aspire to adapt this methodology then to tackle any inputs of unknown lengths.

Key Points for Assessment:

• Does the average accuracy of the produced models significantly exceed 1/128 (given that there are 128 characters in the ASCII set)?
• Does the loss function of the produced models show significant progress within a training epoch?

The results obtained will then serve as the baseline for calculations to determine if this method can effectively reverse engineer SHA-2 algorithms, such as SHA-256, in real-world conditions.

Key results

Work In Progress (WIP)

Usage

Dataset Preparation

To train your custom model, you need a dataset. This dataset is located in the dataset folder and consists of a json file named hash.json.

You have two options to generate those hash.json files :

1. Download Pre-structured Data:

   • Obtain JSON files containing data in an array with two keys: string and hash.

2. Generate Data:

   • Utilize the generate_hash.py script to automatically populate the dataset folder for you. It accepts the following arguments:
     • --num: Number of files to generate (default: 4096)
     • --size: Number of hashes per file (default: 4096)
     • --sufix: If you have generated files before, you can continue the numbering from the next number (default: 1)

You can proceed to the next steps once you have your dataset folder ready with hash JSON files named as follows hash1.json, hash2.json...

Model Setup

For each model or setup, it is recommended to create a dedicated folder. Begin by copying the main folder, then modify the contents of the newly created directory as desired. The main folder includes default parameters and neural layer configurations, along with a pre-trained default model using 16.7 million hashes, which is suitable for testing purposes

Training

Before training your model with train.py, make sure to check and update the constants of your subfolder in the constants.py file to suit your setup.

Pay special attention to these variables:

• DATASET_NB: The number of JSON files in the dataset folder.
• MODEL_PATH: The name of the model file when it will be saved or reused.