Skip to content

README.md

Latest commit

 

History

History
122 lines (85 loc) · 4.18 KB

README.md

File metadata and controls

122 lines (85 loc) · 4.18 KB

LoRA Backdoor Detection from Weights Alone

Repository for the paper:

Detecting Backdoored LoRAs from Weights Alone
Anonymous authors
Under review as a conference paper at COLM 2026 (double-blind)

Overview

This project studies static backdoor detection for LoRA adapters without model execution, trigger search, or training-data access. Given a LoRA adapter, the method reconstructs projection-wise updates in attention (q, k, v, o), extracts spectral/geometric features, and applies a calibrated logistic detector to produce a poison score.

The repository includes:

  • Data generation scripts for benign, poisoned, and test adapter banks.
  • A weight-only detector implementation.
  • Calibration and held-out evaluation scripts.
  • Analysis and plotting scripts for layer/rank sensitivity and cross-model geometry.

Method Summary

For each selected layer and projection, the pipeline computes a compact descriptor from LoRA updates using five statistics:

  • Leading singular value (sigma_1)
  • Frobenius norm (||deltaW||_F)
  • Spectral energy concentration
  • Spectral entropy
  • Kurtosis of flattened update entries

Projection-wise descriptors are concatenated into a 20-dimensional representation and standardized. A logistic regression model maps this representation to a score, and a threshold is selected on validation data (gap-based when strictly separable, otherwise Youden-style).

Repository Structure

.
├── bankCreation/                # Adapter-bank construction scripts
├── core/                        # Detector and feature extraction logic
├── evaluation/                  # Calibration, evaluation, and analysis scripts
├── plotScripts/                 # Figure generation scripts
├── resultsFinal/                # Generated figures/reports (organized by subfolder)
├── config.py                    # Global experiment configuration
└── requirements.txt             # Python dependencies

Environment Setup

python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -r requirements.txt

If required by your workflow, create a .env file with authentication tokens (for example HF_TOKEN).

End-to-End Workflow

1) Build adapter banks

python bankCreation/benignBank.py
python bankCreation/poisonBank.py
python bankCreation/testSet.py
python bankCreation/build_reference_bank.py

2) Calibrate detector

python evaluation/calibrate_detector.py

This step fits the detector and writes calibration artifacts (threshold, reports, distributions).

3) Evaluate on held-out adapters

python evaluation/evaluate_test_set.py

This step writes held-out metrics and error analysis artifacts.

Plot and Analysis Scripts

Current plotting scripts read inputs from plotScripts/... and write outputs to resultsFinal/... by subfolder:

  • plotScripts/layerRankPlots/plot_layer_rank_heatmap.py
  • plotScripts/layerRankPlots/plot_backdoor_detection_correlation.py
  • plotScripts/hackTokenPlots/plot_hack_tokens.py
  • plotScripts/crossModelPlots/cross_model_similarity.py --plots-only

Example:

python plotScripts/layerRankPlots/plot_layer_rank_heatmap.py
python plotScripts/layerRankPlots/plot_backdoor_detection_correlation.py
python plotScripts/hackTokenPlots/plot_hack_tokens.py
python plotScripts/crossModelPlots/cross_model_similarity.py --plots-only

Key Experimental Setting (Paper)

  • Backbones: Llama-3.2-3B-Instruct, Qwen2.5-3B, Gemma-2-2B
  • Detector input: projection-wise spectral/geometric signature from LoRA weights
  • Calibration/test protocol: separate calibration bank and held-out test bank
  • Attack families: rare-token and contextual trigger poisoning
  • Poisoning rates: 1%, 3%, and 5%

For full definitions, equations, and ablations, refer to the submitted manuscript.

Reproducibility Notes

  • Main script behavior is controlled by config.py.
  • Most evaluation scripts assume expected folder layouts under the project root.
  • Generated artifacts (plots, reports, caches) can be large; keep runs isolated if needed.

Citation

This repository is anonymized for double-blind review. Please cite the final camera-ready paper once available.