Lello Molinario
AI Forensics · Digital Forensics · Robustness Evaluation

AI Forensics Research

This page summarizes my research work on the operational robustness of AI-based tools in digital forensics. The focus is not adversarial machine learning in isolation, but the reliability of AI-assisted forensic workflows when image evidence is exposed to realistic distribution shifts, adversarial perturbations, and anti-forensic manipulations.

Final Dataset

1500

Frozen, manually reviewed images divided into weapon, non-weapon, and out-of-distribution classes.

Adversarial Subset

1000

Binary weapon/non-weapon subset used for adversarial and anti-forensic robustness experiments.

Framework

FAIR-Lab

Forensic AI Robustness Lab: a reproducible evaluation framework for AI-based forensic workflows.

Research Objective

The research evaluates whether AI-based image classification systems remain reliable when used in forensic-oriented scenarios affected by image perturbations, source heterogeneity, and operational uncertainty.

The goal is to support forensic decision-making through transparent benchmarking, traceable datasets, robustness metrics, and human-in-the-loop validation.

Dataset Design

The frozen dataset contains three balanced groups: 500 weapon images, 500 non-weapon images, and 500 OOD images. Selection is based on explicit manual criteria and preserved through manifests, hashes, and traceability logs.

weapon: 500non_weapon: 500ood: 500

Evaluation Conditions

  • Clean image classification baseline.
  • Adversarial attacks on the binary subset.
  • Anti-forensic perturbations such as recompression, resizing, blur, histogram modification, and contrast changes.
  • OOD evaluation for borderline or distribution-shifted inputs.
  • Explainability analysis with visual attribution methods.

Models and Tools

The methodology combines transparent proxy models with evaluation of forensic-oriented workflows. The proxy layer is used to quantify robustness, compare conditions, and generate interpretable evidence for the thesis.

EfficientNetResNetCLIPBLIPSVM baselineXAI

Human-in-the-Loop Protocol

The dataset construction process explicitly includes human review. Rather than treating the benchmark as fully automatic, the methodology documents the selection protocol, final labels, excluded samples, and reproducibility constraints. This keeps the final benchmark frozen, auditable, and aligned with the operational nature of digital forensics.

Research Outputs

  • Frozen dataset manifest and adversarial subset manifest.
  • Clean, adversarial, anti-forensic, and OOD evaluation metrics.
  • Comparative robustness tables and confusion matrices.
  • Explainability maps for selected failure cases.
  • Methodological documentation for reproducible forensic AI evaluation.

The public repository is kept minimal and methodologically oriented, with selected scripts and documentation intended to support the thesis rather than expose unnecessary working files.

Professional Context

This research contributes to a broader professional focus on digital forensics, cyber investigation, AI governance, and the reliability of AI-assisted investigative tools.