Beyond Mortality: Advancements in Post-Mortem Iris Recognition through Data Collection and Computer-Aided Forensic Examination

1. Problem Statement

Post-mortem iris recognition offers a critical capability for forensic identification but faces significant barriers:

• Data Scarcity: Collecting biometric data from deceased subjects is logistically difficult, ethically complex, and rare. Existing datasets are small, lack demographic diversity, and rarely include pre-mortem vs. post-mortem comparisons of the same individual.
• Algorithmic Limitations: Most iris recognition algorithms are designed for living subjects. They struggle with post-mortem changes such as corneal clouding, tissue decay, and structural deformation.
• Security Vulnerabilities: There is a risk that post-mortem iris patterns could be used as "presentation attacks" (spoofs) to bypass security systems if the system cannot distinguish between a living and a deceased eye.
• Lack of Forensic Tools: There is a shortage of automated tools that not only match identities but also provide explainable results to support human forensic examiners.

2. Methodology

A. Data Collection (NIJ-2018-DU-BX-0215 Dataset)

The authors collected a new, large-scale dataset from the Dutchess County Medical Examiner's Office (DCMEO):

• Subjects: 259 deceased individuals (518 eyes).
• Imaging: Captured using two sensors:
  • NIR: IriTech IriShield (ISO-compliant near-infrared).
  • Visible Light (RGB): OmniVision OV8865 embedded in a Microsoft Surface tablet.
• Scope: 5,770 NIR and 4,643 RGB images collected across 53 sessions.
• Unique Features:
  • PMI Range: Post-mortem intervals (PMI) ranging from 0.5 hours to 1,674 hours (~70 days).
  • Pre/Post-Mortem Pairing: One subject was imaged both before and after death (a first in published literature).
  • Pathological Cases: Includes cases with dislocated eyeballs and severe cranial trauma where the globe remained intact.
• Curation: Images were anonymized, rotated to align eye canthi, and cropped. Deep learning-based segmentation masks were generated for iris texture, corneal wrinkles, and occlusions.

B. Experimental Design

The study combined the new NIJ dataset with existing Warsaw BioBase datasets (v1.0, v2.0, v3.0) to create a corpus of 338 deceased subjects (7,993 images).

1. Recognition Performance:

   • Evaluated five distinct iris recognition methods:
     • OSIRIS: Classical Gabor-wavelet based.
     • USIT: Non-Gabor kernel based.
     • HDBIF: Human-driven, saliency-based features (trained on post-mortem data).
     • DGR: Deep learning-based Dynamic Graph Representation.
     • VeriEye: Commercial matcher.
   • Analyzed performance across different PMIs, genders, and age groups.

2. Presentation Attack Detection (PAD):

   • Utilized D-NetPAD (a DenseNet-121 based model) to detect if an iris image is from a deceased person (simulating a spoof).
   • Conducted 8 experiments ranging from zero-shot learning (pre-trained model) to few-shot learning (fine-tuning with as few as 5 post-mortem samples) and cross-dataset validation.

3. Demographic Analysis:

   • Balanced datasets by PMI, age, and gender to isolate the impact of these factors on recognition accuracy.

C. Tool Development

• Developed PMExpert, an open-source forensic tool integrating three recognition methods (TLPIM, PBM, HDBIF) with explainability features (visualizing salient regions) and ISO/IEC 29794-6 quality metrics.

3. Key Contributions

1. The NIJ-2018-DU-BX-0215 Dataset: The largest and most diverse public post-mortem iris dataset to date, featuring the longest PMI (1,674 hours) and the first pre-/post-mortem paired samples.
2. Few-Shot PAD Capability: Demonstrated that deep learning models (D-NetPAD) can be fine-tuned with extremely limited data (as few as 5 samples) to effectively distinguish post-mortem irises from living ones, achieving high Area Under Curve (AUC) scores.
3. Forensic Tool (PMExpert): An open-source software suite that combines automated matching with human-interpretable visualizations, currently deployed in a real-world medical examiner's office.
4. Comprehensive Benchmarking: Provided the most extensive evaluation of post-mortem iris recognition using five different algorithmic families on a unified, large-scale dataset.

4. Key Results

Recognition Performance

• Best Performer: HDBIF (Human-Driven Binarized Image Features) consistently outperformed other methods. Its segmentation model was specifically trained on post-mortem data, allowing it to adapt to decay artifacts.
  • Metrics: Equal Error Rate (EER) of ~10-12%, AUC ~96%.
• Worst Performers: DGR (Deep Graph Representation) and USIT showed the poorest performance, likely due to a lack of domain-specific adaptation to post-mortem tissue changes.
• Wavelength Comparison: NIR images significantly outperformed RGB (visible light) images. Cross-wavelength matching (NIR gallery vs. RGB probe) yielded the worst results.
• Pre/Post-Mortem Matching: VeriEye achieved the best results for matching ante-mortem to post-mortem images (EER = 1%), proving the viability of identifying the deceased using living records.

Impact of Demographics and PMI

• PMI (Time since death): The most significant degradation in recognition occurs within the first 72 hours. After 72 hours, the rate of performance decline slows, suggesting that identity information is lost rapidly initially but stabilizes.
• Gender: Statistical tests (ANOVA/Kruskal-Wallis) indicated significant performance differences between genders, though no single matcher consistently favored one gender across all metrics.
• Age: Subjects aged 34–66 generally showed better recognition performance than younger (1–33) or older (67–99) groups for most matchers.

Presentation Attack Detection (PAD)

• Fine-tuning Effectiveness: A pre-trained D-NetPAD model failed to detect post-mortem irises without fine-tuning (AUC ~0.57).
• Few-Shot Learning: Fine-tuning with just 50 samples yielded AUCs of 0.995–0.997. Even with only 5 samples, the model maintained reasonable discriminative ability (AUC ~0.93), proving that forensic labs can build effective detectors with minimal data.
• Correlation with PMI: Higher PMI generally increased the likelihood of successful detection (easier to distinguish from living tissue).

5. Significance

• Forensic Utility: This work validates that iris recognition is a viable tool for identifying deceased individuals, even weeks after death, provided the eyes are preserved (e.g., in cold storage).
• Security Implications: The study highlights a critical vulnerability: standard iris systems may accept post-mortem irises as valid. However, it also provides a solution via few-shot PAD models that can be deployed to prevent such spoofing.
• Resource Availability: By releasing the NIJ-2018-DU-BX-0215 dataset and the PMExpert tool via the National Archive of Criminal Justice Data (NACJD), the authors have democratized access to high-quality forensic data, enabling further research and standardization in the field.
• Human-in-the-Loop: The introduction of explainable AI (visualizing why a match occurred) bridges the gap between automated algorithms and the legal requirements of forensic testimony, making the technology more acceptable for court use.