Fractal Characterization of Low-Correlation Signals in AI-Generated Image Detection

Abstract: AI-generated imagery has reached near-photorealistic fidelity, yet this technology poses significant threats to information security and societal trust. Existing deepfake detection methods often exhibit limited robustness in open-world scenarios. To address this limitation, this paper investigates intrinsic discrepancies between synthetic and authentic images from a signal-level perspective. Our analysis reveals that low-correlation signals serve as distinctive markers for differentiating AI-generated imagery from real photographs. Building on this insight, we introduce a novel method for quantifying these signals based on fractal theory. By analyzing the fractal characteristics of low-correlation signals, our method effectively captures the subtle statistical anomalies inherent to the synthesis process. Extensive experimental results demonstrate the method's robustness and superior detection performance. This work emphasizes the need to shift research focus to a new signal-level direction for deepfake detection. Theoretically, this proposed approach is not limited to face image identification but can be applied to all AI-generated image detection tasks. This study provides a new research direction for deepfake detection.

• The paper demonstrates that isolating low-correlation signals via PCA reveals subtle yet robust statistical differences between genuine and AI-generated images.
• It employs fractal measures like fractal dimension, multifractal spectrum, and Shannon entropy on PCA residuals, achieving KS statistics over 0.93 for effective discrimination.
• This approach generalizes well against adversarial tactics and novel generative models, providing a scalable solution for synthetic image forensics.

Fractal Characterization of Low-Correlation Signals for AI-Generated Image Detection

Introduction and Motivation

The proliferation of high-fidelity AI-generated imagery, powered by generative models such as GANs and diffusion models, has precipitated critical challenges for authenticity assessment in digital media. Legacy detection paradigms, which predominantly exploit discriminative local texture and artifact cues via CNN-based architectures, display substantial performance degradation and limited generalization when confronted with novel generative techniques and diverse real-world scenarios. Transformer-based vision models and frequency-domain discriminators have only partially mitigated these limitations, underscoring the necessity for fundamentally new perspectives. This study frames deepfake detection from a signal-level analytic viewpoint, isolating low-correlation image signals—components orthogonal or weakly correlated to main object semantics—as robust discriminative markers.

Theoretical Framework

The authors introduce a taxonomy that distinguishes image signals based on their correlation with primary semantic content. High-correlation signals, closely associated with the salient objects in an image, are largely mimicked by state-of-the-art generative models. In contrast, low-correlation signals, such as fine-grained background textures and stochastic noise patterns, reflect statistical properties inadequately modeled by current generation pipelines. These low-correlation signals, when isolated, expose subtle yet consistent discrepancies between authentic and AI-generated content.

Principal Component Analysis (PCA) provides the operational backbone for separating high- and low-correlation signals. By projecting face-cropped images onto their principal components, the approach reconstructs images using only a specified subset of components. The residuals—the difference between the input and this reconstruction—encode low-correlation information while suppressing semantic signals tied to the main object.

Fractal theory is employed to capture the geometric and statistical complexity of these residuals. Specifically, fractal dimension (FD), lacunarity, multifractal spectrum (MFS), and Shannon entropy serve as feature primitives. These metrics quantify microstructural textural deviations, heterogeneity, and information-theoretical properties, elucidating the generative footprints left by synthesis models.

Methodology

The experimental protocol is executed on large-scale face datasets encompassing both authentic (FFHQ) and synthetic (1 Million Fake Faces) samples. The pipeline consists of the following:

1. Preprocessing: MTCNN is used for facial region extraction.
2. Signal Separation: PCA decomposes and reconstructs the cropped face regions. Residual images are constructed by excluding the top $N$ principal components, where $N$ is varied to sweep the boundary between high- and low-correlation signals.
3. Feature Extraction: The aforementioned fractal and statistical features are computed for both the original and residual images.
4. Statistical Validation: The Kolmogorov-Smirnov (KS) test quantifies the separability of the feature distributions between real and synthetic classes in both raw and residual domains.

Experimental Results and Analysis

Raw face images exhibit no visually apparent or statistically significant separation in the distributions of entropy, FD, MFS, or lacunarity. This is reflected in high p-values from normality tests, and the overlapping visualizations of corresponding distributions across both classes.

In marked contrast, PCA suppression of high-correlation components (with $N$ between 24 and 32) results in residuals where the calculated features—FD, entropy, MFS, and standard deviation—demonstrate pronounced and statistically robust inter-class differences. For example, KS statistics for FD and entropy consistently exceed 0.93 with near-zero $p$-values across all tested residuals, and standard deviation displays maximal discrimination. The MFS-based metrics show increasingly clear class separation as more high-correlation components are suppressed, confirming the efficacy of low-correlation signals as discriminators.

Lacunarity, however, remains statistically non-discriminative in the residual domain, suggesting that while it is informative of texture inhomogeneity, it does not capture the generative artifacts exploited by the other fractal measures in this context.

Implications

This work provides empirical and theoretical evidence that focusing on low-correlation signals—effectively those orthogonal or weakly related to depicted semantics—is significantly more robust for AI-generated image detection than conventional holistic or main-object-focused feature sets. Fractal analytic techniques, particularly FD and MFS, offer a principled avenue for quantifying generative anomalies, suggesting broad applicability to both face and non-face generative image detection.

In practical terms, this approach circumvents overfitting to dataset-specific texture or artifact biases, generalizes effectively to new generative methods, and is not constrained by adversarial attacks that target high-level semantic content. The extracted fractal fingerprints from low-correlation signals may serve as generative model signatures, enabling forensic analyses, source attribution, and reliable detection in real-world, open-set conditions.

Future Directions

Potential extensions include the application of this pipeline to multi-modal and non-face image synthesis domains, integration with frequency-domain feature learners, and exploration of adaptive selection for the optimal number of principal components suppressed. The fingerprinting potential of low-correlation fractal features for generative model attribution warrants systematic investigation, particularly in adversarial and online scenarios. Further, learning-based or hybrid pipelines may incorporate fractal features as explicit priors or bottlenecks to enhance detector robustness.

Conclusion

Fractal analysis of low-correlation signals, isolated via PCA-based suppression of main-object semantics, is demonstrated to be a robust and generalizable approach for distinguishing AI-generated images from real images. The statistical and textural irregularities uncovered in the low-correlation residual domain serve as effective discriminators, validated both qualitatively and quantitatively. This methodology offers a well-founded, domain-agnostic foundation for future research in synthetic media detection and forensic analysis.