Implementing a KNN Image Classifier from Scratch: Practice in Distinguishing AI-Generated and Real Images

This project implements the K-Nearest Neighbors (KNN) algorithm from scratch, builds a high-precision AI-generated image recognition system through systematic feature engineering experiments and grid search cross-validation, solves the problem of distinguishing AI-generated images from real ones, and fully demonstrates the standard workflow of a machine learning project.

Generative AI technology is developing rapidly, and the quality of AI-generated images is approaching real photos, bringing challenges to content authenticity verification. This project explores technical solutions for distinguishing AI-synthesized images from real photos by building a machine learning classification system.

The core goal is to build a binary classification system (labeling AI-generated/real photos). KNN is implemented from scratch to ensure transparent and controllable workflow. The Kaggle "AI vs Real Images" dataset is used, covering multiple categories; undersampling is adopted to balance classes (250 images per class, 400 in training set, 100 in test set) to ensure the model is not affected by distribution skew.

Four feature extraction methods are compared: 1. RGB flattening (3072 dimensions, retains spatial information but has high dimensionality and is sensitive); 2. Grayscale flattening (1024 dimensions, reduces dimensionality but loses color); 3. Color histogram (focuses on global color statistics, spatially invariant); 4. Grayscale histogram (texture and brightness analysis). Experimental results: Histogram-based features are better than pixel flattening, and color histogram is the optimal solution.

The core components of KNN are implemented from scratch: supports Euclidean/Manhattan distance; the prediction process includes distance calculation, selecting k neighbors, and majority voting. Testing k values found k=9 to be optimal. 5-fold cross-validation + grid search (covering 4 features, 6 k values, 2 distances, 3 histogram bins) is used, and the optimal configuration is: color histogram (32 bins), 64x64 size, k=9, Manhattan distance.

Independent test set results: overall accuracy 84%, Macro F1 0.8399; F1 for AI-generated images is 0.8431 (43/50 correct), F1 for real images is 0.8367 (41/50 correct). Compared to the baseline (RGB flattening + KNN) with 48% accuracy and 0.3404 F1, the improvement is significant, verifying the value of feature engineering and tuning.

Highlights: 1. Educational value (fully demonstrates the supervised learning workflow); 2. Evaluation metrics (Macro F1 considers class balance); 3. Rigorous experiments (cross-validation avoids data leakage). Applications: content moderation, news authenticity, copyright protection. The methodology can be migrated to complex deep learning models.

This project demonstrates the standard workflow of machine learning. Core insight: In the era of deep learning, traditional ML methods and rigorous feature engineering still have value, suitable for resource-constrained or interpretability-required scenarios. The project code and dataset are open-sourced, suitable for ML entry-level practice.