This project addresses two primary tasks related to image classification and adversarial robustness, along with explainability of AI-generated images. The repository is structured into three main folders: Task1, Attacks, and Task2.
This folder contains implementations of various methods to classify an image as real or fake.
- Multiple classification algorithms to identify AI-generated images.
- Configurable pipelines for training, evaluation, and testing.
- Comprehensive metrics to evaluate model performance, including accuracy, precision, and recall.
This folder focuses on adversarial attacks designed to test and inhibit the robustness of models developed in Task1. The attacks aim to expose vulnerabilities in classification models by manipulating input images.
- Scripts for generating adversarial examples using various attack strategies.
This folder deals with explaining why certain images are classified as fake. The explanation process leverages Vision-Language Models (VLMs) fine-tuned on resized images to provide insights into the artifacts and patterns that led to classification.
- Scripts for fine-tuning VLMs on task-specific data.
- Inference pipelines for generating explainability outputs.
- Tools to visualize and interpret artifacts present in fake images.
- Refer to the individual folders (
task1,Attacks,task2) for detailed implementation documentation, scripts, and configuration options. - Each folder contains a
README.mdfile with specific instructions for its respective code.