Integration of Certified Defense Methods (Randomized Smoothing)

[henba1]

Integration of Certified Defense Methods (Randomized Smoothing) - Addresses #85

Overview

This PR implements the first iteration of probabilistic certified defense methods, specifically Randomized Smoothing, into VERONA as outlined in issue #85. The implementation extends VERONA's EpsilonValueEstimator and VerificationModule abstract classes to support certification with configurable sampling rates, smoothing levels, and confidence parameters.

Changes Made

New Components

1. ProbabilisticCertificationResult (robustness_experiment_box/database/verification_result.py)

• New data structure for probabilistic certification results
• Includes to_dict() and from_dict() methods for serialization
• Stores certified radius, predicted class, and confidence information

2. RandomizedSmoothingModule (robustness_experiment_box/verification_module/randomized_smoothing_module.py)

• Extends VerificationModule abstract base class
• Implements verify_probabilistic() → returns ProbabilisticCertificationResult
• Core method _perform_randomized_smoothing() orchestrates the certification process:
  • _sample_noise() - generates Gaussian noise samples
  • _apply_diffusion_denoising() - placeholder for DDS integration (TODO)
  • _lower_confidence_bound() - computes statistical confidence bounds
• Configurable parameters: σ (noise level), N₀ (prediction samples), N (certification samples), α (confidence)

3. RandomizedSmoothingEstimator (robustness_experiment_box/epsilon_value_estimator/randomized_smoothing_estimator.py)

• Extends EpsilonValueEstimator abstract base class
• compute_epsilon_value() - bridges to classical EpsilonValueResult format
• get_probabilistic_result() - returns ProbabilisticCertificationResult using RandomizedSmoothingModule
• Maintains compatibility with existing VERONA experiment framework

4. Extended VerificationModule Interface (robustness_experiment_box/verification_module/verification_module.py)

• Added new abstract method to base VerificationModule class
• Enables probabilistic verification across all verification modules

5. Example Script (scripts/randomized_smoothing_example.py)

• Demonstrates usage of the new randomized smoothing functionality
• Shows integration with VERONA's robustness experiment framework
• Provides practical example of parameter configuration

Implementation Details

This implementation follows Option A: Full Integration as discussed in #85, providing:

• end-to-end certification pipeline with Monte Carlo sampling
• Direct integration with VERONA's existing architecture
• Backward compatibility - no changes needed to epsilon_value_result.py
• Framework preparation for diffusion-enhanced smoothing (DDS placeholder implemented)

Algorithm Support

Randomized Smoothing (Cohen et al.)

• Monte Carlo sampling with Gaussian noise
• Statistical bounds for probabilistic certification
• Configurable sampling parameters (σ, N₀, N, α)

Diffusion Denoised Smoothing (Carlini et al.)

• Framework prepared with _apply_diffusion_denoising() placeholder
• TBD DDS integration

Notes on Integration with VERONA

• integration with existing EpsilonValueEstimator and VerificationModule interfaces
• No breaking changes to existing result structures

Testing

• Added comprehensive example script demonstrating the new functionality
• Integration with existing VERONA robustness experiment framework
• Maintains backward compatibility with existing verification modules

Future Work

This is the first iteration of the Randomized Smoothing integration. Future iterations include:

• Implementation of Diffusion Denoised Smoothing (DDS)
• Performance optimizations

References

• Cohen et al. - Certified Adversarial Robustness via Randomized Smoothing
• Carlini et al. - (Certified!!) Adversarial Robustness for Free!

Files Changed

• robustness_experiment_box/epsilon_value_estimator/randomized_smoothing_estimator.py (new)
• robustness_experiment_box/verification_module/randomized_smoothing_module.py (new)
• robustness_experiment_box/verification_module/verification_module.py (extended)
• robustness_experiment_box/database/verification_result.py (new)
• scripts/randomized_smoothing_example.py (new)

---

Addresses #85 (First iteration - issue remains open for future updates)