ARQ: A Mixed-Precision Quantization Framework for Accurate and Certifiably Robust DNNs

2024-10-31

arXiv.org

AI Summary

This work addresses the challenge of simultaneously achieving high accuracy and provable robustness in mixed-precision quantized deep neural networks. We propose the first quantization framework that integrates reinforcement learning (RL) with randomized smoothing. Methodologically, we employ RL to automatically search layer-wise mixed-precision quantization policies, jointly optimized via a robustness-aware loss and certified through randomized smoothing to maximize certified radius while preserving accuracy. Experiments demonstrate that our quantized models retain full floating-point baseline accuracy, with only 1.5% instruction overhead over the original model. Across multiple datasets and adversarial perturbations, our approach achieves superior certified robust accuracy and certified radii compared to state-of-the-art methods. To the best of our knowledge, this is the first method to jointly guarantee accuracy, empirical robustness, and formal certification under low computational overhead constraints.

Technology Category

• Machine Learning: Adversarial Learning & Robustness
• Computer Vision: Adversarial Attacks & Robustness
• Natural Language Processing: Safety and Robustness

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

Mixed precision quantization has become an important technique for optimizing the execution of deep neural networks (DNNs). Certified robustness, which provides provable guarantees about a model's ability to withstand different adversarial perturbations, has rarely been addressed in quantization due to unacceptably high cost of certifying robustness. This paper introduces ARQ, an innovative mixed-precision quantization method that not only preserves the clean accuracy of the smoothed classifiers but also maintains their certified robustness. ARQ uses reinforcement learning to find accurate and robust DNN quantization, while efficiently leveraging randomized smoothing, a popular class of statistical DNN verification algorithms. ARQ consistently performs better than multiple state-of-the-art quantization techniques across all the benchmarks and the input perturbation levels. The performance of ARQ quantized networks reaches that of the original DNN with floating-point weights, but with only 1.5% instructions and the highest certified radius. ARQ code is available at https://anonymous.4open.science/r/ARQ-FE4B.

Problem

• Optimizes deep neural networks via mixed-precision quantization
• Ensures certified robustness against adversarial perturbations
• Reduces computational cost while maintaining high accuracy

Innovation

• Mixed-precision quantization for DNNs
• Reinforcement learning for robustness
• Randomized smoothing for verification