Lei Hsiung

selected publications

1. 

   Why LLM Safety Guardrails Collapse After Fine-tuning: A Similarity Analysis Between Alignment and Fine-tuning Datasets

   Lei Hsiung, Tianyu Pang, Yung-Chen Tang, Linyue Song, Tsung-Yi Ho, Pin-Yu Chen, and Yaoqing Yang

   In Submission, 2025

   Abs arXiv Demo

   Recent advancements in large language models (LLMs) have underscored their vulnerability to safety alignment jailbreaks, particularly when subjected to downstream fine-tuning. However, existing mitigation strategies primarily focus on reactively addressing jailbreak incidents after safety guardrails have been compromised, removing harmful gradients during fine-tuning, or continuously reinforcing safety alignment throughout fine-tuning. As such, they tend to overlook a critical upstream factor: the role of the original safety-alignment data. This paper therefore investigates the degradation of safety guardrails through the lens of representation similarity between upstream alignment datasets and downstream fine-tuning tasks. Our experiments demonstrate that high similarity between these datasets significantly weakens safety guardrails, making models more susceptible to jailbreaks. Conversely, low similarity between these two types of datasets yields substantially more robust models and thus reduces harmfulness score by up to 10.33%. By highlighting the importance of upstream dataset design in the building of durable safety guardrails and reducing real-world vulnerability to jailbreak attacks, these findings offer actionable insights for fine-tuning service providers.

2. 

   Spectral Insights into Data-Oblivious Critical Layers in Large Language Models

   Xuyuan Liu, Lei Hsiung, Yaoqing Yang, and Yujun Yan

   In Findings of the Association for Computational Linguistics (Findings of ACL), 2025

   Abs arXiv Demo

   Understanding how feature representations evolve across layers in large language models (LLMs) is crucial for enhancing their interpretability and robustness. Recent research has explored this by identifying critical layers associated with specific functions or behaviors in LLMs. However, prior work has primarily identified critical layers through data-dependent analyses of fine-tuned models, limiting their applicability to post-hoc scenarios. In contrast, we propose a data-oblivious method to pinpoint intrinsic critical layers in pre-fine-tuned LLMs by analyzing representation dynamics using Centered Kernel Alignment (CKA). We demonstrate that layers exhibiting significant changes in representation space correspond to those undergoing the most substantial modifications during fine-tuning, and we show that this phenomenon remains consistent across tasks for a given model. Furthermore, our novel spectral analysis reveals that shifts in layer representations are driven by changes in their top principal components, which govern semantic transformations from rationales to conclusions. Notably, freezing these layers during fine-tuning reduces adversarial attack success rates by 40%, enabling efficient, minimally invasive defenses. Our findings suggest that critical layers are an intrinsic property of pre-fine-tuned models, providing a unified framework for understanding LLM layer dynamics and facilitating resource-efficient adaptation strategies.

3. 

   AutoVP: An Automated Visual Prompting Framework and Benchmark

   Hsi-Ai Tsao*, Lei Hsiung*, Pin-Yu Chen, Sijia Liu, and Tsung-Yi Ho (* Equal Contribution)

   In Proceedings of the Twelfth International Conference on Learning Representations (ICLR), 2024

   Abs arXiv Code

   Visual prompting (VP) is an emerging parameter-efficient fine-tuning approach to adapting pre-trained vision models to solve various downstream image-classification tasks. However, there has hitherto been little systematic study of the design space of VP and no clear benchmark for evaluating its performance. To bridge this gap, we propose AutoVP, an end-to-end expandable framework for automating VP design choices, along with 12 downstream image-classification tasks that can serve as a holistic VP-performance benchmark. Our design space covers 1) the joint optimization of the prompts; 2) the selection of pre-trained models, including image classifiers and text-image encoders; and 3) model output mapping strategies, including nonparametric and trainable label mapping. Our extensive experimental results show that AutoVP outperforms the best-known current VP methods by a substantial margin, having up to 6.7% improvement in accuracy; and attains a maximum performance increase of 27.5% compared to linear-probing (LP) baseline. AutoVP thus makes a two-fold contribution: serving both as an efficient tool for hyperparameter tuning on VP design choices, and as a comprehensive benchmark that can reasonably be expected to accelerate VP’s development.

4. 

   NeuralFuse: Learning to Recover the Accuracy of Access-Limited Neural Network Inference in Low-Voltage Regimes

   Hao-Lun Sun, Lei Hsiung, Nandhini Chandramoorthy, Pin-Yu Chen, and Tsung-Yi Ho

   In Proceedings of the 38th Annual Conference on Neural Information Processing Systems (NeurIPS), 2024

   Abs arXiv Code

   Deep neural networks (DNNs) have become ubiquitous in machine learning, but their energy consumption remains problematically high. An effective strategy for reducing such consumption is supply-voltage reduction, but if done too aggressively, it can lead to accuracy degradation. This is due to random bit-flips in static random access memory (SRAM), where model parameters are stored. To address this challenge, we have developed NeuralFuse, a novel add-on module that handles the energy-accuracy tradeoff in low-voltage regimes by learning input transformations and using them to generate error-resistant data representations, thereby protecting DNN accuracy in both nominal and low-voltage scenarios. As well as being easy to implement, NeuralFuse can be readily applied to DNNs with limited access, such cloud-based APIs that are accessed remotely or non-configurable hardware. Our experimental results demonstrate that, at a 1% bit-error rate, NeuralFuse can reduce SRAM access energy by up to 24% while recovering accuracy by up to 57%. To the best of our knowledge, this is the first approach to addressing low-voltage-induced bit errors that requires no model retraining.

5. 

   Towards Compositional Adversarial Robustness: Generalizing Adversarial Training to Composite Semantic Perturbations

   Lei Hsiung, Yun-Yun Tsai, Pin-Yu Chen, and Tsung-Yi Ho

   In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023

   Abs arXiv Code Demo

   Model robustness against adversarial examples of single perturbation type such as the \(\ell_p\)-norm has been widely studied, yet its generalization to more realistic scenarios involving multiple semantic perturbations and their composition remains largely unexplored. In this paper, we first propose a novel method for generating composite adversarial examples. Our method can find the optimal attack composition by utilizing component-wise projected gradient descent and automatic attack-order scheduling. We then propose generalized adversarial training (GAT) to extend model robustness from \(\ell_p\)-ball to composite semantic perturbations, such as the combination of Hue, Saturation, Brightness, Contrast, and Rotation. Results obtained using ImageNet and CIFAR-10 datasets indicate that GAT can be robust not only to all the tested types of a single attack, but also to any combination of such attacks. GAT also outperforms baseline \(\ell_∞\)-norm bounded adversarial training approaches by a significant margin.