Hongyi Li

Hongyi Li, Jun Xu, William Ward Armstrong

We introduce the Learning Hyperplane Tree (LHT), a novel oblique decision tree model designed for expressive and interpretable classification. LHT fundamentally distinguishes itself through a non-iterative, statistically-driven approach to constructing splitting hyperplanes. Unlike methods that rely on iterative optimization or heuristics, LHT directly computes the hyperplane parameters, which are derived from feature weights based on the differences in feature expectations between classes within each node. This deterministic mechanism enables a direct and well-defined hyperplane construction process. Predictions leverage a unique piecewise linear membership function within leaf nodes, obtained via local least-squares fitting. We formally analyze the convergence of the LHT splitting process, ensuring that each split yields meaningful, non-empty partitions. Furthermore, we establish that the time complexity for building an LHT up to depth $d$ is $O(mnd)$, demonstrating the practical feasibility of constructing trees with powerful oblique splits using this methodology. The explicit feature weighting at each split provides inherent interpretability. Experimental results on benchmark datasets demonstrate LHT's competitive accuracy, positioning it as a practical, theoretically grounded, and interpretable alternative in the landscape of tree-based models. The implementation of the proposed method is available at https://github.com/Hongyi-Li-sz/LHT_model.

Hongyi Li, Jiawei Ye, Jie Wu et al.

Large Language Models (LLMs) aligned with human feedback have recently garnered significant attention. However, it remains vulnerable to jailbreak attacks, where adversaries manipulate prompts to induce harmful outputs. Exploring jailbreak attacks enables us to investigate the vulnerabilities of LLMs and further guides us in enhancing their security. Unfortunately, existing techniques mainly rely on handcrafted templates or generated-based optimization, posing challenges in scalability, efficiency and universality. To address these issues, we present JailPO, a novel black-box jailbreak framework to examine LLM alignment. For scalability and universality, JailPO meticulously trains attack models to automatically generate covert jailbreak prompts. Furthermore, we introduce a preference optimization-based attack method to enhance the jailbreak effectiveness, thereby improving efficiency. To analyze model vulnerabilities, we provide three flexible jailbreak patterns. Extensive experiments demonstrate that JailPO not only automates the attack process while maintaining effectiveness but also exhibits superior performance in efficiency, universality, and robustness against defenses compared to baselines. Additionally, our analysis of the three JailPO patterns reveals that attacks based on complex templates exhibit higher attack strength, whereas covert question transformations elicit riskier responses and are more likely to bypass defense mechanisms.

James Flemings, Haosheng Gan, Hongyi Li et al.

In-context learning (ICL) has shown promising improvement in downstream task adaptation of LLMs by augmenting prompts with relevant input-output examples (demonstrations). However, the ICL demonstrations can contain privacy-sensitive information, which can be leaked and/or regurgitated by the LLM output. Differential Privacy (DP), a widely adopted privacy safeguard, has emerged to mitigate this privacy leakage, with recent work demonstrating strong privacy-utility tradeoffs in classification tasks for ICL. However, generation tasks for ICL are challenging due to the high-dimensional output space of open-ended generation. To this end, we propose $\texttt{dps-mozo}$, Differentially Private Sampling by Mixing One-shot with Zero-shot Outputs, a decoding framework that generates DP text by sampling from the product of multiple one-shot outputs mixed with a zero-shot output. This mixing effectively reduces the amount of information that can be leaked by each demonstration. By utilizing the inherent randomness in sampling from the mixed distributions, we can achieve DP without adding noise, thereby improving the privacy-utility tradeoff. Our experimental evaluations show $\texttt{dps-mozo}$ can achieve a strong privacy guarantee, $ε=2$, with minimal utility degradation compared to non-private few-shot learning, $\textbf{0.3}$% ROUGE-L F1 score decrease on the SAMSum dataset with Gemma 2 2B.

Hongyi Li, Jun Xu, William Ward Armstrong

This paper introduces a novel tree-based model, Learning Hyperplane Tree (LHT), which outperforms state-of-the-art (SOTA) tree models for classification tasks on several public datasets. The structure of LHT is simple and efficient: it partitions the data using several hyperplanes to progressively distinguish between target and non-target class samples. Although the separation is not perfect at each stage, LHT effectively improves the distinction through successive partitions. During testing, a sample is classified by evaluating the hyperplanes defined in the branching blocks and traversing down the tree until it reaches the corresponding leaf block. The class of the test sample is then determined using the piecewise linear membership function defined in the leaf blocks, which is derived through least-squares fitting and fuzzy logic. LHT is highly transparent and interpretable--at each branching block, the contribution of each feature to the classification can be clearly observed.