PoisonSpot

Nov 22, 2025·
Philemon Hailemariam
Philemon Hailemariam
Birhanu Eshete
Birhanu Eshete
· 1 min read
projects

Project Snapshot

Item Details
Paper PoisonSpot: Precise Spotting of Clean-Label Backdoors via Fine-Grained Training Provenance Tracking
Venue ACM CCS 2025
Primary Theme Clean-label backdoor defense
Main Artifacts Zenodo code/data package + paper

Authors

  • Philemon Hailemariam
  • Birhanu Eshete

Overview

PoisonSpot focuses on precise detection of clean-label backdoor attacks by tracking how individual training samples affect parameter updates over time. The system attributes poisoning influence through provenance lineage and flags high-risk samples for rejection.

What This Project Delivers

  • Fine-grained training provenance capture inspired by dynamic taint tracking.
  • Sample-level poisoning attribution to detect stealthy clean-label attacks.
  • Artifact package for reproducible benchmarking under adaptive threat models.

Repository and Paper

Trustworthy ML Adversarial ML Backdoor Detection Training Provenance
Philemon Hailemariam
Authors
Philemon Hailemariam
PhD Candidate
Birhanu Eshete
Authors
Birhanu Eshete
Principal Investigator
Associate Professor of Computer Science at the University of Michigan-Dearborn and Director of the Data-Driven Security & Privacy Lab (DSPLab).