💻 computer science

Score-based Likelihood Ratios for Deepfake Image Evidence Using Deep Learning Features

This paper proposes a forensically compliant, score-based likelihood ratio framework that leverages SRM deep learning features and calibrated statistical modeling to quantify the evidential strength of deepfake images, overcoming the limitations of conventional binary detection by providing interpretable, numerically grounded assessments for judicial decision-making.

Original authors: Tianli Guo, Pengji Chen, Rui Lu, Yunqi Tang

Published 2026-09-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Tianli Guo, Pengji Chen, Rui Lu, Yunqi Tang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the courtroom, the weight of evidence is rarely a simple matter of "yes" or "no." When a piece of digital media is presented as proof, the question is not just whether it is real or fake, but how strongly the data supports one conclusion over the other. For decades, forensic experts have relied on analyzing lighting, shadows, and digital fingerprints to spot tampering. However, a new generation of artificial intelligence has changed the game. These systems can now swap faces in videos and photos with such precision that the results look indistinguishable from reality to the human eye, even to trained experts. Traditional tools that look for obvious errors or missing data often fail because these new forgeries are generated to be perfect. This creates a critical gap: the legal system needs a way to measure the strength of evidence when the image itself looks flawless, moving beyond a binary guess to a calculated probability.

To bridge this gap, researchers Tianli Guo, Pengji Chen, Rui Lu, and Yunqi Tang have developed a new method for evaluating deepfake images that treats the problem as a statistical question rather than a simple classification task. Instead of training a computer to shout "fake" or "real," their system measures how closely a questioned image resembles a collection of known, genuine faces. They start by using a specialized neural network, a type of artificial intelligence designed to spot subtle, high-frequency patterns that human eyes cannot see, to extract a unique digital fingerprint from every image. These fingerprints are not just simple numbers but complex maps of texture and structure. The researchers then calculate the average fingerprint of all the genuine images they have collected, creating a central point of reference, or a "centroid," that represents the ideal of authenticity.

The core of their work involves measuring the distance between a new, unknown image and this central point of genuine faces. If an image is truly real, its digital fingerprint should sit very close to the center of the genuine group. If it is a deepfake, even a high-quality one, its fingerprint will drift away from that center. The researchers tested this approach by comparing thousands of real and fake images, mapping out exactly how these distances are distributed. They found that while some forgeries are very good at mimicking reality, they still leave a statistical trace that pushes them further from the center of the genuine group than a real photo would. By using a mathematical technique called kernel density estimation, they turned these distances into a likelihood ratio. This ratio acts as a scale, telling a forensic examiner exactly how much more likely it is that the image is real compared to the chance that it is a forgery, rather than just offering a vague confidence score.

To ensure this system is reliable enough for a courtroom, the team applied several layers of rigorous testing and correction. They used a method known as the Pool Adjacent Violators algorithm to calibrate their results, ensuring that the numbers they produced accurately reflected the true probability of the evidence. Their tests on a specific dataset of deepfake images showed that the system could distinguish between real and fake with a high degree of accuracy, making errors in only about 6.75% of cases. Furthermore, they addressed a common problem in statistical modeling where extreme values can lead to misleading conclusions. By setting empirical boundaries, they ensured that the system never produces a result that is less reliable than a random guess, effectively cutting off the most extreme and uncertain predictions. This process, which they call the Empirical Lower and Upper Bound method, trimmed away less than 9% of the most extreme data points, leaving a robust and trustworthy set of results.

The findings demonstrate that it is possible to move beyond the limitations of current deepfake detectors, which often rely on fixed thresholds that can be easily fooled. By combining deep learning features with careful statistical modeling, the researchers have created a pipeline that produces a quantifiable, interpretable measure of evidence strength. This approach does not claim to be a magic bullet that instantly solves every case, but it provides a forensically compliant framework that aligns with the standards of legal evidence. The system successfully translates the complex, invisible differences between real and fake images into a clear, numerical value that can be understood and weighed in a judicial setting. While the current study focuses on face-swapping deepfakes within specific datasets, the method offers a promising path forward for a legal system that needs to navigate an increasingly complex digital landscape with precision and confidence.

Technical Summary: Score-based Likelihood Ratios for Deepfake Image Evidence Using Deep Learning Features

Problem Statement
The rapid advancement of deepfake technology, particularly in face-swapping using Generative Adversarial Networks (GANs) and diffusion models, has rendered traditional forensic authentication methods ineffective. Conventional approaches rely on binary classification with fixed thresholds (e.g., 0.5), which lack probabilistic interpretation and cannot quantify the graded strength of evidence required in legal proceedings. Furthermore, existing deep learning detectors often produce uncalibrated confidence scores that are sensitive to data distribution shifts and fail to distinguish between ambiguous cases and clear forgeries. There is a critical need for a framework that transforms "black-box" AI outputs into transparent, numerically grounded evidence compliant with Bayesian inference standards used in forensic science (e.g., DNA and fingerprint analysis).

Methodology
The authors propose a Score-based Likelihood Ratio (SLR) framework driven by deep learning features to address these forensic gaps. The methodology consists of four primary stages:

Deep Feature Extraction:
- The study utilizes the Spatial Rich Model (SRM) network, a dual-branch architecture designed to jointly analyze RGB content and high-frequency frequency-domain features.
- The network employs an Xception backbone with five stages, progressively reducing spatial resolution while increasing channel depth to 2048.
- A cross-modal attention mechanism fuses the branches, outputting 2048-channel, 8×8 spatial feature maps. These are flattened into 1D vectors for statistical analysis.
Centroid-Referenced Similarity Scoring:
- Instead of pairwise random matching, the study constructs a global genuine feature centroid ( $\mu_0$ ) derived from the mean feature vectors of all authentic training samples.
- Similarity scores are calculated using Cosine Similarity between individual test samples (both real and deepfake) and this universal real-class centroid.
- This generates two score distributions: intra-authentic (real vs. real centroid) and cross-forgery (fake vs. real centroid).
Probability Density Estimation and Calibration:
- Kernel Density Estimation (KDE) with a Gaussian kernel is applied to model the probability density functions (PDFs) of the similarity scores for both hypotheses ( $H_p$ : authentic; $H_d$ : deepfake).
- Bandwidths are determined using Silverman's rule and optimized via ten-fold cross-validation.
- The Likelihood Ratio (LR) is computed as the ratio of the probability densities: $LR = f(Score | H_p) / f(Score | H_d)$ .
- Pool Adjacent Violators (PAV) algorithm is used to calibrate the scores, ensuring they are well-calibrated probabilistic measures.
Empirical Bounds (ELUB):
- To prevent misleading evidence from tail-region extrapolation, the Empirical Lower and Upper Bound (ELUB) method is applied.
- Using the Normalized Bayes Error-rate (NBE) curve with one Consequential Misleading LR (CMLR), the system identifies a reliable operating range.
- LR values outside this range are truncated to the nearest boundary, ensuring the system never performs worse than a neutral baseline ( $LR=1$ ).

Key Contributions

Forensic-Compliant Framework: The study bridges the gap between deep learning detection and judicial standards by converting similarity scores into interpretable likelihood ratios, moving beyond binary "real/fake" classifications.
Universal Centroid Paradigm: The authors introduce a unified genuine feature centroid approach, allowing for standardized similarity measurement across diverse datasets and overcoming the instability of random pairwise matching.
Robust Calibration and Bounding: The integration of PAV calibration and ELUB truncation ensures that the output LRs are statistically reliable, well-calibrated, and resistant to over-extrapolation in extreme cases.
Quantifiable Evidence: The pipeline provides a numerically grounded assessment of evidential strength, supporting judicial decisions with transparent statistical backing.

Results
The framework was evaluated on four mainstream datasets (FaceForensics++, Celeb-DF-v2, DFDCP, UADFV), with in-depth analysis focused on the DFDCP dataset due to its superior performance in preliminary evaluations.

Discrimination: The system achieved an Equal Error Rate (EER) of 6.75% on the DFDCP test set.
Calibration: The Cost of Log-Likelihood Ratio (Cllr) was reduced from 0.3306 (uncalibrated) to 0.3174 after PAV calibration, indicating improved reliability and reduced evidential risk.
Distribution Analysis: DET curves, Tippett plots, and Expected Calibration Error (ECE) analyses confirmed strong discrimination between hypotheses and stable calibration across different prior probabilities.
ELUB Impact: The ELUB method constrained log10LR values to the range [−1.796, 2.303]. This truncation affected only 8.88% of samples (1,044 out of 11,764), effectively eliminating unreliable tail extrapolations while preserving the core discriminative distribution.

Significance and Claims
The paper claims to provide a forensically compliant, quantifiable pipeline for deepfake evidence assessment. By uniting deep feature learning (SRM) with calibrated statistical modeling (KDE, PAV, ELUB), the study delivers evidence strength that is:

Interpretable: Transformed from opaque AI scores into transparent likelihood ratios.
Numerically Grounded: Based on rigorous statistical distributions rather than arbitrary thresholds.
Judicially Admissible: Designed to align with the Bayesian inference framework widely adopted in forensic science, offering a practical solution for evaluating high-fidelity face-swapped forgeries where human visual judgment is insufficient.

The authors acknowledge limitations, noting that the evaluation is currently restricted to face-swapped deepfakes on specific datasets and relies solely on the SRM network without multimodal augmentation. Future work is proposed to expand to diverse deepfake types and validate the ELUB bounds across different datasets.

Technical Summary: Score-based Likelihood Ratios for Deepfake Image Evidence Using Deep Learning Features

More like this