Backdoor Detection

This research project focuses on detecting backdoor attacks in the PreActResNet architecture under data-limited configurations within a 1-minute time limit (tested on an NVIDIA RTX 4090).

Problem Statement

The complete problem description is available in the problem_statement.pdf.

Explorations

Various ideas and methodologies explored during this research are documented in the analysis directory. Before running the notebooks, ensure they are moved to the project's root directory.

Usage

The main detection method is implemented in main.py with the following function:

def backdoor_model_detector(model: PreActResNet, 
                            num_classes: int,
                            test_images_folder_address: str,
                            transformation: transforms.Compose
                           ) -> bool:
    """
    Detects whether the given model is malicious (contains a backdoor) or not.
    """
    # Function implementation

This function analyzes the provided PreActResNet model to determine if it has been compromised by a backdoor attack.

Related Works

Interesting articles and novel ideas are stored in the related_works directory.

Results

This model was developed for the Rayan Trustworthy AI Challenge, achieving a score of 65/100 and ranking 12th.