SynthDetect Ultra: AI Image Forensics

SynthDetect Ultra is an advanced forensic tool for detecting AI-generated (synthetic) images using a combination of physics-inspired, statistical, and deep forensic features. It is built with Streamlit for an interactive web-based analysis experience.

Features

• Gradient Field Analysis: Measures anisotropy and coherence in image gradients to detect unnatural patterns.
• Frequency Domain Analysis: Analyzes spectral slope and high-frequency ratios to spot GAN/Diffusion artifacts.
• Benford’s Law (DCT): Checks for naturalness in DCT coefficient distributions.
• CFA Artifacts: Examines color filter array correlations typical of real camera sensors.
• Texture Analysis (GLCM): Extracts homogeneity and contrast from gray-level co-occurrence matrices.
• Chromatic Aberration: Detects lens-induced color fringing absent in AI images.
• Noise Residual Kurtosis: Measures the “tailedness” of noise, distinguishing real sensor noise from AI noise.
• ELA (Error Level Analysis): Assesses JPEG compression error patterns for inconsistencies.
• Grid Artifact Detection: Finds periodic patterns in the frequency domain typical of upsampling artifacts.
• Noise Consistency: Compares noise correlation across RGB channels.
• Color Space Anomalies: Analyzes LAB and HSV distributions for unnatural color statistics.
• Micro-Texture Analysis: Uses wavelet transforms to examine fine texture kurtosis and symmetry.
• Edge Profile Analysis: Evaluates edge sharpness, consistency, and uniformity.
• EXIF Metadata Forensics: Flags suspicious or missing metadata, including AI software tags.
• Ensemble Decision Engine: Weighted scoring system for robust, explainable verdicts.
• Detection History: Stores and summarizes past analyses in a local SQLite database.

Tags: Forensics, Explainable AI, Image Analysis, Python, Streamlit

Usage

1. Install Requirements

   pip install -r requirements.txt
   

   Ensure you have the following key packages: streamlit, opencv-python, numpy, scipy, matplotlib, pandas, scikit-image, pywt, Pillow.

2. Run the App

   streamlit run app_1.py
   

3. Upload an Image

   • Supported formats: JPG, PNG, JPEG, WEBP
   • The app will process the image and display a detailed forensic report.

4. Interpret Results

   • The dashboard provides a confidence score, core and advanced metrics, and visualizations.
   • Flags highlight suspicious features (e.g., lack of lens aberration, artificial noise, grid artifacts).
   • A radar chart and summary statistics offer a multi-dimensional authenticity profile.

Forensic Metrics and Equations

1. Gradient Field Analysis

• Anisotropy (ρ): \(\rho = \frac{\lambda_1}{\lambda_2}\) where $\lambda_1, \lambda_2$ are eigenvalues of the gradient covariance matrix.
• Coherence (κ): \(\kappa = \left(\frac{\lambda_1 - \lambda_2}{\lambda_1 + \lambda_2}\right)^2\)

2. Frequency Domain Analysis

• Spectral Slope (β): \(\beta = -\text{slope of linear fit to } \log_{10}(\text{radial frequency}) \text{ vs. } \log_{10}(\text{power})\)
• High-Frequency Ratio (η): \(\eta = \frac{\sum_{f>f_c} P(f)}{\sum_{f>0} P(f)}\) where $P(f)$ is the radial power spectrum, $f_c$ is a cutoff frequency.

3. Benford’s Law (DCT)

• Benford Divergence: \(D = \sum_{d=1}^9 \frac{(p_d - b_d)^2}{b_d}\) where $p_d$ is the observed proportion of first digits, $b_d = \log_{10}(1 + 1/d)$ is the Benford distribution.

4. CFA Artifacts

• CFA Score: \(ext{CFA} = \text{corr}(G - R, G - B)\)

5. Texture Analysis (GLCM)

• Homogeneity: \(\sum_{i,j} \frac{P(i,j)}{1 + |i-j|}\)
• Contrast: \(\sum_{i,j} (i-j)^2 P(i,j)\) where $P(i,j)$ is the normalized GLCM.

6. Chromatic Aberration

• Radial Shift Ratio: \(ext{CA} = \frac{\text{mean}|G-R|_{\text{edge}}}{\text{mean}|G-R|_{\text{center}}}\)

7. Noise Residual Kurtosis

• Kurtosis: \(ext{Kurtosis} = \frac{E[(X - \mu)^4]}{\sigma^4}\) where $X$ is the noise residual, $\mu$ is the mean, $\sigma$ is the standard deviation.

8. ELA (Error Level Analysis)

• ELA Mean: \(ext{ELA Mean} = \text{mean}(|I_{\text{orig}} - I_{\text{jpeg}}|)\)

9. Grid Artifact Score

• Grid Score: \(ext{Grid Score} = \frac{\text{# horizontal peaks} + \text{# vertical peaks}}{2}\)

10. Noise Consistency

• Noise Correlation: \(ext{Noise Corr} = \text{mean}([\text{corr}(R,G), \text{corr}(R,B), \text{corr}(G,B)])\)

11. Color Space Anomalies

• Color Anomaly Score: \(ext{Color Anomaly} = \frac{|\text{skew}(a^*)| + |\text{skew}(b^*)|}{2}\)

12. Micro-Texture (Wavelet Kurtosis)

• Wavelet Kurtosis: \(ext{mean}(|\text{kurtosis of detail coefficients}|)\)

13. Edge Profile Analysis

• Edge Consistency: \(ext{Edge Consistency} = \frac{\text{std}(\text{edge gradients})}{\text{mean}(\text{edge gradients})}\)

Example Testing Results

Replace the above image with your actual results screenshot or visualization.

Customization

• Adjust detection thresholds in the sidebar for sensitivity tuning.
• View detection history and statistics in the sidebar.

Limitations

• Some heavily post-processed or compressed real images may trigger false positives.
• The tool is designed for single-image analysis; batch or video support is not included.

Citation

If you use SynthDetect Ultra in your research or project, please cite appropriately.

Author: Koshik Debanath

License: MIT

Tech Stack: Python, Streamlit

Links:

• Try Demo
• Watch Demo