Training efficiency often increases model vulnerability to attacks
No Free Efficiency: Revisiting the Trade-off Between Training Efficiency and Model Vulnerability
Machine LearningCryptography and Security
Summary
Training big AI models efficiently is important but might make them less safe. The authors found that techniques to speed up training make models easier to trick or attack. They studied this problem across image and language AI systems and found consistent weaknesses linked to how the models internally work. They suggest future AI training should balance speed, performance, and safety all at once.
What this means in practice
- •For ai engineering teams: Adjust training pipelines to monitor and reduce increased vulnerability caused by efficiency-focused training methods.
- •For security analysts: Use insights about efficiency-related weaknesses to better anticipate and defend against attacks on AI models.
Authors
Yiyong Liu, Jun Sakuma, Michael Backes, Rui Wen
Abstract
Training efficiency has become the central driver of recent progress in foundation models. To overcome the massive computational and data requirements of large-scale training, researchers increasingly adopt strategies such as selective data sampling, efficient pre-training, and simplified reinforcement learning pipelines. While these strategies drastically reduce overhead, they prompt a critical, yet neglected question: Is efficiency achieved at the expense of model robustness and security? To our knowledge, we present the first systematic cross-domain investigation of the efficiency-vulnerability trade-off. Across vision and language models, we show that efficiency-oriented training increases susceptibility to adversarial and privacy attacks. We characterize this vulnerability by analyzing the models' internal geometry and functional representations, demonstrating that the evaluated efficient variants consistently exhibit sharper loss geometry together with systematic changes in representational structure. We further extend our analysis to "zero RL training", finding that models trained using simplified RL recipes exhibit substantially greater susceptibility to catastrophic forgetting and more pronounced overconfidence than those trained through conventional alignment pipelines. Our findings suggest that training efficiency is rarely a "free lunch"; rather, the mechanisms that minimize computation can inadvertently compromise safety. We conclude by calling for a paradigm shift toward multi-objective training that jointly optimizes for performance, cost, and security.