This repository contains the implementation for our study focusing on two primary objectives:
- Experimental validation of Projected Gradient Descent (PGD) adversarial training in improving neural networks' intrinsic stability against adversarial attacks.
- Demonstration of leveraging stability-enhanced models for advanced backdoor detection using an Adversarially Enhanced Trigger Inversion method.
Core Hypothesis: Models trained with adversarial robustness exhibit more stable behavior—both in classification accuracy under attack and internal feature representation consistency. This stability enhances higher-level security analyses like backdoor trigger inversion.
- Dataset: CIFAR-10
- Path:
./data/cifar10
generate_test_triggers.py: Creates test triggers including:- Backdoored triggers (trained via a poisoned SimpleEncoder model).
- Benign triggers (random noise, zero tensors, or weights extracted from clean models).
-
Control Group (Standard Training):
python train.py --train_mode standard --num_models 5 --output_dir ./evaluation_models --epochs_per_encoder 20
-
Experimental Group (PGD Adversarial Training):
python train.py --train_mode pgd --num_models 5 --output_dir ./evaluation_models --epsilon_pgd 0.03 --steps_pgd 10 --step_size_pgd 0.01 --epochs_per_encoder 20
Saved model weights will be stored in ./evaluation_models.
Run the evaluation script:
python run_stability_evaluation.pyNote: Due to computational constraints, the default setup uses a limited number of test samples. For higher statistical reliability, consider increasing the sample size in a more powerful environment.
Key Findings:
- PGD adversarial training significantly improves model robustness against adversarial attacks in the DECREE framework.
- Enhanced stability in feature embeddings enables more reliable inversion and detection of backdoor triggers.
Our experiments demonstrate that:
- PGD adversarial training effectively boosts encoder robustness in DECREE-based scenarios.
- Stability-enhanced models provide a stronger foundation for security-critical tasks like backdoor detection.
This README provides a structured, concise overview of your project. Replace `results.png` with your actual figure path, and expand sections like "Results" with specific metrics/plots as needed.