Efficient Adversarial Training via Criticality-Aware Fine-Tuning

Researchers introduce Criticality-Aware Adversarial Training (CAAT), a parameter-efficient method that identifies and fine-tunes only the most robustness-critical parameters in Vision Transformers, achieving 94.3% of standard adversarial training robustness while tuning just 6% of model parameters. This breakthrough addresses the computational bottleneck preventing large-scale adversarial training deployment.

The research tackles a fundamental scalability challenge in machine learning security. Vision Transformers have revolutionized computer vision by scaling effectively to massive datasets, yet their adversarial robustness fails to improve proportionally with model size. Traditional adversarial training requires full-model fine-tuning, making it computationally prohibitive for large architectures—a barrier that has limited deployment in production systems. CAAT solves this by strategically identifying which parameters most influence robustness, then applying parameter-efficient fine-tuning only to critical modules. This approach represents a meaningful shift from brute-force robustness enhancement toward intelligent resource allocation. The technical contribution matters because adversarial robustness remains essential for deploying AI systems in safety-critical domains like autonomous vehicles, medical imaging, and security applications. The minimal robustness loss (4.3%) against dramatic parameter reduction (94% fewer trained parameters) demonstrates the method's efficiency gains are real, not theoretical. This work opens pathways for making adversarial training practical at enterprise scale, where computational budgets are finite. The broader implication extends beyond vision tasks—similar criticality-aware approaches could optimize fine-tuning for other deep learning domains facing similar computational constraints. As AI systems increasingly influence high-stakes decisions, methods enabling scalable robustness become infrastructure rather than academic curiosity.

• →CAAT achieves 94.3% of full adversarial training robustness while fine-tuning only 6% of Vision Transformer parameters, dramatically reducing computational costs.
• →The method identifies robustness-critical parameters adaptively, enabling targeted parameter-efficient fine-tuning rather than full-model updates.
• →Experiments across three adversarial learning datasets demonstrate CAAT outperforms existing lightweight adversarial training methods.
• →This approach addresses the critical gap between ViT scalability and adversarial robustness, enabling practical large-scale deployment.
• →Parameter-efficient fine-tuning techniques can be strategically applied to robustness problems, not just domain adaptation tasks.