Adversarial Attack Optimization & Cross-Architecture Robustness
Multi-dimensional optimization framework for efficient adversarial attacks
Project Overview
Developed a multi-dimensional optimization framework for adversarial attacks, integrating gradient decoupling, hybrid noise strategies, and model ensemble approaches to improve attack efficiency and cross-architecture transferability.
Core Algorithms
FastJSMA (Fast Jacobian-based Saliency Map Attack)
- Gradient Decoupling: Optimized traditional JSMA complexity to 1/80 of original method
- Performance: Achieved 81x speedup on CIFAR-100 while maintaining 98%+ success rate
- Scalability: First method to support ImageNet large-scale attacks, breaking memory limitations
Season (Spectrum-Aware Orthogonal Gradient Refinement)
- Spectrum-aware decoupling: Gaussian-smoothed low-frequency vs. high-frequency surrogate-gradient branches for structure vs. texture
- Low-saliency guidance: Precomputed saliency on the clean image; steers high-frequency updates toward background while preserving foreground shape cues for ViT-like targets
- Orthogonal refinement: Projects accumulated high-frequency perturbation onto the orthogonal complement of the low-frequency structural component to reduce branch interference
- Results: Training-free wrapper over standard transfer attacks; on ImageNet (ResNet-50 surrogate, unified protocol) average transfer success gains of +6.6 percentage points (up to +16.0) across CNN, ViT, and MLP targets vs.\ strong baselines
- Venue: ICME 2026 (IEEE International Conference on Multimedia and Expo, CCF-B), accepted March 17, 2026, Spotlight presentation; Shengjie Xu, corresponding author — publication entry
Watertox (Lightweight Universal Attack Framework)
- Two-stage FGSM: Implemented multi-model ensemble with gradient voting
- Efficiency: Achieved 1 image/second processing speed
- Performance: Reduced target model accuracy by up to 98.8% in zero-shot attacks
- Competitive Edge: Outperformed contemporary methods by 3%-5%
Technical Achievements
Algorithm Innovation
- Efficiency Breakthrough: Solved efficiency and compatibility bottlenecks in large-scale scenarios
- Cross-Architecture Support: Successfully tested on 20+ model architectures
- Dataset Coverage: Comprehensive evaluation on CIFAR, ImageNet, and other major datasets
Research Impact
- ICCV 2025: FastJSMA accepted (co-first author); presented in Honolulu
- ICME 2026: Season accepted March 17, 2026 (CCF-B), Spotlight presentation; corresponding author — View Publication
- Publications: Watertox on arXiv (2024) — View Publication
- Industry Application: Methods applicable to model security and robustness testing
Key Contributions
Technical Leadership
- Algorithm Design: Led core algorithm development and experimental validation
- Cross-Method Optimization: Coordinated multiple attack strategies for enhanced performance
- Comprehensive Testing: Covered multiple datasets and model architectures
Research Output
- Publications: FastJSMA at ICCV 2025 (co-first author); Season at ICME 2026 (corresponding author, Spotlight); Watertox preprint (2024)
- Performance: Strong results on attack efficiency (FastJSMA) and cross-architecture transfer (Season)
- Practical Impact: Methods suitable for security evaluation of heterogeneous vision systems