DeepProv
Project Snapshot
| Item | Details |
|---|---|
| Paper | DeepProv: Behavioral Characterization and Repair of Neural Networks via Inference Provenance Graph Analysis |
| Venue | ACSAC 2025 |
| Primary Theme | Trustworthy model behavior analysis |
| Main Artifacts | Code + artifact dataset |
Authors
- Firas Ben Hmida
- Abderrahmen Amich
- Ata Kaboudi
- Birhanu Eshete
Overview
DeepProv introduces inference provenance graphs (IPGs) to characterize neural network behavior during inference and guide systematic repair. It supports empirical and structural analysis to improve robustness and reliability.
What This Project Delivers
- Provenance-oriented representation of model behavior during inference.
- Node/edge-level repair strategies for robustness-oriented hardening.
- Artifacts for experiments across models, attacks, and evaluation settings.
Repository and Paper
- Code: github.com/um-dsp/DeepProv
- Data Artifact: github.com/um-dsp/DeepProv/tree/main/artifact/data
- Paper: arxiv.org/pdf/2509.26562
