SPECTRA: Semantic Perturbation-Based Counterfactuals and Training for Robustness Against Adversarial Attacks

Overview

SPECTRA (Semantic Perturbation-based Counterfactuals for Robust Adversarial Training) is a novel framework that enhances interpretability and robustness in deep learning models. It generates counterfactuals by applying minimal perturbations in the embedding space of CNNs, focusing on fine-grained semantic attributes. This project utilizes the CUB-200 dataset and demonstrates improved model stability against adversarial attacks.

Methodology

1. Dataset

   • Uses the CUB-200 dataset with 312 semantic attributes across 200 bird species.
   • Enables fine-grained, interpretable perturbations.

2. Model Architecture

   • Implemented models include a linear classifier, simple CNN, complex CNN, and ResNet.
   • Attribute vector embeddings aid in semantic counterfactual generation.

3. Counterfactual Generation

   • Perturbations applied in embedding space to achieve minimal feature changes.
   • Uses optimization techniques to generate meaningful attribute shifts.

4. Robustness Evaluation

   • Introduces the SPECTRA Attackability Metric, measuring perturbation resistance.
   • Evaluates adversarial stability by computing minimal norm perturbations.

Key Features

• Semantic Counterfactuals: Generates interpretable counterfactuals based on attribute shifts.
• Robust Adversarial Training: Improves model resilience against adversarial perturbations.
• Minimal Perturbation Calculation: Ensures targeted and localized feature changes.
• Attackability Metric: Provides a quantitative measure of model robustness.

Results

Quantitative Metrics:

• Attack Success Rate (ASR)
• Perturbation Noise Ratio
• Peak Signal-to-Noise Ratio (PSNR)

Qualitative Analysis:

• Localized attribute modifications
• Visualization of perturbed features

Acknowledgments

This project was developed by Aditya Sridhar and Ananya Varshney as part of research on adversarial robustness in deep learning.

Contact

For inquiries, please reach out to:

• Aditya Sridhar - aditya.sridharr.11@gmail.com
• GitHub: My GitHub profile