Authorship Identifier

Java project for identifying whether C/C++ source code was written by a human or by LLMs AI.

The project workflow has three steps:

1. Extract static code attributes from source files.
2. Train a multilayer perceptron network using sigmoid activation function and backpropagation.
3. Validate all files in the validate folder using the trained model.

Project Structure

src/
  CodeAttributeExtractor.java   # Shared extraction logic for the 50 attributes
  ExtractCodeAttributes.java    # Generates training CSV files in trainingData
  MultiLayerPerceptron.java     # Trains the MLP and saves the model in models
  ValidateFile.java             # Validates all files in the validate folder

trainingCodebase/
  human/                        # Human-written source code used for training
  ai/                           # AI-written source code used for training

trainingData/                   # Generated training CSV files
models/                         # Trained models saved as .txt files
validate/                       # Files to classify

Requirements

• Java JDK 26, or a compatible JDK for this project.

1. Extract Code Attributes

ExtractCodeAttributes walks through an input folder, filters C/C++ files (.c, .h, .cpp, .hpp, .cc, .cxx, .hh, .hxx), and generates a CSV with 50 attributes per file.

Arguments

• <output_file.csv>: name or path of the CSV file that will receive the extracted attributes.
• <input_folder>: optional folder containing the source files to analyze.

Argument Rules

• If only <output_file.csv> is provided, the program uses trainingCodebase as the input folder.
• If <output_file.csv> is only a file name, the CSV is saved inside trainingData.
• If <input_folder> is human or ai, the program automatically looks inside trainingCodebase.
• If <input_folder> is a relative or absolute path, that folder is used directly.
• If no argument is provided, the program prints the expected argument format and stops.
• The generated CSV starts with a file column, followed by the 50 extracted attributes in a fixed order.

2. Train the MLP

MultiLayerPerceptron loads two CSV files: one for the HUMAN class and one for the AI class.

• Class 0 = HUMAN
• Class 1 = AI

The network is a multilayer perceptron implemented in Java. It uses an input layer with one neuron for each extracted attribute, one hidden layer, and a single output neuron.

Each neuron uses the sigmoid activation function:

sigmoid(x) = 1 / (1 + e^(-x))

The output is interpreted as a probability-like score:

• values closer to 0 indicate HUMAN;
• values closer to 1 indicate AI.

Training is performed with backpropagation. For each sample, the network performs a forward pass, calculates the output error, propagates that error back through the hidden layer, and updates weights and biases using the learning rate. Before training, the input attributes are normalized using the mean and standard deviation calculated from the training split. The final model stores the normalization data, all weights, and all biases in a text file.

Arguments

• <human_csv>: CSV file containing features extracted from human-written code.
• <ai_csv>: CSV file containing features extracted from AI-written code.
• <model_file.txt>: optional output file for the trained model.

Argument Rules

• <human_csv> and <ai_csv> are required.
• If <human_csv> or <ai_csv> is only a file name, the program looks for it inside trainingData.
• If <human_csv> or <ai_csv> is a relative or absolute path, that file is used directly.
• If <model_file.txt> is not provided, the model is saved as models/model_mlp.txt.
• If <model_file.txt> is only a file name, it is saved inside models.
• If fewer than two arguments are provided, the program prints the expected argument format and stops.

3. Validate Files

ValidateFile validates every C/C++ file found in the validate folder.

The program reads the .txt model file and automatically infers:

• the number of inputs from the size of MEANS or MEDIAS;
• the number of hidden neurons from the size of W1, B1, and W2.

Validation uses the same CodeAttributeExtractor class used during training data generation, ensuring that both training and prediction use the exact same feature extraction logic.

Arguments

• <model_file.txt>: optional model file to load.
• <validate_folder>: optional folder containing the source files to classify.

Argument Rules

• If no argument is provided, the program loads models/model_mlp.txt and validates the validate folder.
• If <model_file.txt> is provided as a simple file name, the program looks for it inside models.
• If <model_file.txt> is provided as a relative or absolute path, that model file is used directly.
• If <validate_folder> is not provided, the program validates the validate folder.
• If <validate_folder> is provided as a simple folder name, the program first checks that folder directly and can also resolve it inside validate.
• If <validate_folder> is provided as a relative or absolute path, that folder is used directly.
• If more than two arguments are provided, the program prints the expected argument format and stops.
• If the model file does not exist, validation stops.
• If the validation folder does not exist or contains no C/C++ files, validation stops.

4. Extracted Attributes

The project uses 50 attributes for each source file.

Structural Attributes

1. Number of classes
2. Number of functions
3. Number of methods
4. Number of imports/includes
5. Number of function calls
6. Number of returns
7. Number of assignments
8. Number of variables
9. Number of function parameters
10. Number of decorators/attributes
11. Number of lambdas
12. Number of templates
13. Number of try/catch blocks
14. Number of throws
15. Number of asserts

Control-Flow Attributes

16. Number of if statements
17. Number of else statements
18. Number of else if statements
19. Number of for loops
20. Number of while loops
21. Number of break statements
22. Number of continue statements
23. Number of boolean operators
24. Number of comparisons
25. Number of arithmetic operators

Complexity Attributes

26. Average cyclomatic complexity
27. Maximum cyclomatic complexity
28. Maximum block depth
29. Average block depth
30. Average function size
31. Maximum function size
32. Average number of statements per function
33. Average number of branches per function

Size and Density Attributes

34. Total number of lines
35. Number of blank lines
36. Number of comment lines
37. Comment-to-line ratio
38. Average line length
39. Maximum line length
40. Average number of characters per function

Identifier and Style Attributes

41. Number of unique identifiers
42. Total number of identifiers
43. Unique identifier ratio
44. Average identifier length
45. Maximum identifier length
46. Number of snake_case identifiers
47. Number of camelCase identifiers
48. snake_case ratio
49. camelCase ratio
50. Identifier entropy

All attributes are extracted by CodeAttributeExtractor, which is shared by ExtractCodeAttributes and ValidateFile.