HO-FMN: Hyperparameter Optimization for Fast Minimum-Norm Attacks

2024-07-11

Neurocomputing

AI Summary

Existing gradient-based adversarial attack methods—relying on fixed loss functions, optimizers, and hyperparameters—often overestimate model robustness and suffer from inefficient minimal-norm perturbation search and high hyperparameter sensitivity. To address these limitations, this work introduces Bayesian hyperparameter optimization into the Fast Minimum-Norm (FMN) attack framework for the first time, establishing an end-to-end automated tuning pipeline that jointly optimizes the attack objective, step-size scheduling, and gradient update mechanism. Evaluated on CIFAR-10, CIFAR-100, and ImageNet, our method achieves a 3.2× speedup over baselines including PGD and AutoAttack, while reducing average L₂ perturbation magnitude by 18%, significantly improving the trade-off between attack efficiency and perturbation minimization. The core innovation lies in deeply embedding Bayesian optimization within the inner loop of adversarial attack generation, enabling adaptive and fine-grained calibration of robustness evaluation.

Technology Category

Machine Learning: Adversarial Learning & RobustnessComputer Vision: Adversarial Attacks & RobustnessSearch and Optimization: Adversarial Search

Application Category

User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systemsSearch and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for rankingGraph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs

Problem

• Optimizing hyperparameters for minimum-norm adversarial attacks
• Addressing overly-optimistic robustness evaluations in ML models
• Providing dynamic adjustment of attack parameters and loss functions

Innovation

• Dynamic hyperparameter optimization for adversarial attacks
• Parametric variation of fast minimum-norm algorithm
• Adjustable loss, optimizer, and scheduler components