Handcrafted Feature Fusion for Reliable Detection of AI-Generated Images

The rapid progress of generative models has enabled the creation of highly realistic synthetic images, raising concerns about authenticity and trust in digital media. Detecting such fake content reliably is an urgent challenge. While deep learning approaches dominate current literature, handcrafted features remain attractive for their interpretability, efficiency, and generalizability. In this paper, we conduct a systematic evaluation of handcrafted descriptors, including raw pixels, color histograms, Discrete Cosine Transform (DCT), Histogram of Oriented Gradients (HOG), Local Binary Patterns (LBP), Gray-Level Co-occurrence Matrix (GLCM), and wavelet features, on the CIFAKE dataset of real versus synthetic images. Using 50,000 training and 10,000 test samples, we benchmark seven classifiers ranging from Logistic Regression to advanced gradient-boosted ensembles (LightGBM, XGBoost, CatBoost). Results demonstrate that LightGBM consistently outperforms alternatives, achieving PR-AUC 0.9879, ROC-AUC 0.9878, F1 0.9447, and a Brier score of 0.0414 with mixed features, representing strong gains in calibration and discrimination over simpler descriptors. Across three configurations (baseline, advanced, mixed), performance improves monotonically, confirming that combining diverse handcrafted features yields substantial benefit. These findings highlight the continued relevance of carefully engineered features and ensemble learning for detecting synthetic images, particularly in contexts where interpretability and computational efficiency are critical.

💡 Research Summary

The paper investigates whether classic, handcrafted image descriptors can still compete with modern deep‑learning approaches for detecting AI‑generated (synthetic) images. Using the CIFAKE benchmark—a balanced collection of 32 × 32 real and synthetic photographs—the authors extract seven families of features: raw RGB pixel values, per‑channel color histograms, low‑frequency Discrete Cosine Transform (DCT) coefficients, Histogram of Oriented Gradients (HOG), Uniform Local Binary Patterns (LBP), Gray‑Level Co‑occurrence Matrices (GLCM), and first‑level Daubechies wavelet sub‑band statistics. These are evaluated in three configurations: (i) baseline (raw + hist + DCT), (ii) advanced (HOG + LBP + GLCM + wavelets), and (iii) mixed (concatenation of all seven).

Seven classifiers are benchmarked: Logistic Regression, Random Forest, ExtraTrees, HistGradientBoosting, XGBoost, LightGBM, and CatBoost, plus a soft‑voting ensemble of the probabilistic models. All models share comparable hyper‑parameters (500‑1000 trees for the boosting methods, learning rate 0.05, etc.). The training set contains 50 k images, a held‑out validation set (10 % of training) is used to tune the decision threshold that maximizes the F1 score, and performance is finally measured on a 10 k test set.

Metrics include threshold‑free PR‑AUC and ROC‑AUC, as well as threshold‑dependent F1, Matthews Correlation Coefficient (MCC), balanced accuracy, and the Brier score for probability calibration. The results show a clear hierarchy: as feature richness increases, every classifier improves, and among them, gradient‑boosted tree ensembles dominate. LightGBM consistently achieves the highest scores. With baseline features it reaches PR‑AUC 0.9676, ROC‑AUC 0.9666, F1 0.9018, and Brier 0.0712. With advanced features the scores rise to PR‑AUC 0.9852, ROC‑AUC 0.9850, F1 0.9384, Brier 0.0531. The mixed configuration yields the best overall performance: PR‑AUC 0.9879, ROC‑AUC 0.9878, F1 0.9447, MCC 0.8022, balanced accuracy 0.9011, and a low Brier 0.0414, indicating both excellent discrimination and well‑calibrated probabilities.

The authors interpret these findings as evidence that (1) handcrafted descriptors capture complementary cues—global color statistics, frequency‑domain irregularities, edge/shape patterns, and fine‑grained texture—that together expose artifacts left by generative models; (2) tree‑based boosting algorithms are especially adept at handling high‑dimensional, heterogeneous feature vectors, automatically selecting the most informative dimensions while avoiding over‑fitting; and (3) the resulting pipeline is computationally lightweight (feature extraction runs on CPU in milliseconds) and fully interpretable, making it attractive for resource‑constrained forensic settings or policy environments where transparency is required.

Limitations are acknowledged: CIFAKE’s low resolution may not reflect performance on high‑resolution real‑world images; the synthetic images stem from a limited set of generative models, so generalization to newer diffusion models remains untested; and feature engineering is manual, lacking automated selection or dimensionality‑reduction techniques. Future work is proposed to test on larger, higher‑resolution datasets, explore meta‑learning for feature selection, and combine handcrafted descriptors with deep‑learning embeddings in hybrid models.

In conclusion, the study demonstrates that a well‑designed handcrafted feature suite, when paired with modern gradient‑boosted tree ensembles—particularly LightGBM—can achieve detection performance comparable to state‑of‑the‑art deep networks while offering superior interpretability and efficiency. This positions handcrafted‑based pipelines as a viable, practical alternative for AI‑generated image detection, especially in contexts where computational resources are limited or model transparency is paramount.