arXiv subjects

Kuntai Cai

Publications and source records attributed to Kuntai Cai.

5 recordsLinked to original sources

arXiv2026

Inverting the Shield: Systematically Generating Safety Tests from Policy Specifications

The widespread integration of Large Language Models (LLMs) necessitates rigorous and systematic safety evaluation. Existing paradigms either rely on constructed benchmarks to assess safety from predefined perspectives, or employ dynamic red-teaming to probe potential vulnerabilities. While effective, these approaches face challenges, as they depend heavily on expert domain knowledge, offer limited systematic guarantees, and are vulnerable to rapid obsolescence. To address these limitations, we introduce a novel framework POLARIS that brings the rigor of specification-based software testing to AI safety. POLARIS first compiles unstructured natural-language policies into First-Order Logic (FOL) representations, establishing a traceable link between high-level rules and concrete test cases. This formalization enables the construction of a Semantic Policy Graph, where complex policy violation scenarios are encoded as traversable paths. By systematically exploring this graph, POLARIS uncovers compositional violation patterns, which are then instantiated into executable natural-language test queries, enabling coverage-driven and reproducible safety testing. Experiments demonstrate that POLARIS achieves higher policy coverage and attack success counts compared to established baselines. Crucially, by bridging formal methods and AI safety, POLARIS provides a principled, automated approach to ensuring LLMs adhere to safety-critical policies with verifiable traceability. We release our code at https://github.com/huac-lxy/POLARIS.

cs.AI↗

arXiv2025

AutoContext: Instance-Level Context Learning for LLM Agents

Current LLM agents typically lack instance-level context, which comprises concrete facts such as environment structure, system configurations, and local mechanics. Consequently, existing methods are forced to intertwine exploration with task execution. This coupling leads to redundant interactions and fragile decision-making, as agents must repeatedly rediscover the same information for every new task. To address this, we introduce AutoContext, a method that decouples exploration from task solving. AutoContext performs a systematic, one-off exploration to construct a reusable knowledge graph for each environment instance. This structured context allows off-the-shelf agents to access necessary facts directly, eliminating redundant exploration. Experiments across TextWorld, ALFWorld, Crafter, and InterCode-Bash demonstrate substantial gains: for example, the success rate of a ReAct agent on TextWorld improves from 37% to 95%, highlighting the critical role of structured instance context in efficient agentic systems.

cs.CL↗

arXiv2025

PrivPetal: Relational Data Synthesis via Permutation Relations

Releasing relational databases while preserving privacy is an important research problem with numerous applications. A canonical approach is to generate synthetic data under differential privacy (DP), which provides a strong, rigorous privacy guarantee. The problem is particularly challenging when the data involve not only entities (e.g., represented by records in tables) but also relationships (represented by foreign-key references), since if we generate random records for each entity independently, the resulting synthetic data usually fail to exhibit realistic relationships. The current state of the art, PrivLava, addresses this issue by generating random join key attributes through a sophisticated expectation-maximization (EM) algorithm. This method, however, is rather costly in terms of privacy budget consumption, due to the numerous EM iterations needed to retain high data utility. Consequently, the privacy cost of PrivLava can be prohibitive for some real-world scenarios. We present a sophisticated PrivPetal approach that addresses the above issues via a novel concept: permutation relation, which is constructed as a surrogate to synthesize the flattened relation, avoiding the generation of a high-dimensional relation directly. The synthesis is done using a refined Markov random field mechanism, backed by fine-grained privacy analysis. Extensive experiments using multiple real datasets and the TPC-H benchmark demonstrate that PrivPetal significantly outperforms existing methods in terms of aggregate query accuracy on the synthetic data.

cs.DB↗

arXiv2024

GCON: Differentially Private Graph Convolutional Network via Objective Perturbation

Graph Convolutional Networks (GCNs) are a popular machine learning model with a wide range of applications in graph analytics, including healthcare, transportation, and finance. However, a GCN trained without privacy protection measures may memorize private interpersonal relationships in the training data through its model parameters. This poses a substantial risk of compromising privacy through link attacks, potentially leading to violations of privacy regulations such as GDPR. To defend against such attacks, a promising approach is to train the GCN with differential privacy (DP), a rigorous framework that provides strong privacy protection by injecting random noise into the training process. However, training a GCN under DP is a highly challenging task. Existing solutions either perturb the graph topology or inject randomness into the graph convolution operations, or overestimate the amount of noise required, resulting in severe distortions of the network's message aggregation and, thus, poor model utility. Motivated by this, we propose GCON, a novel and effective solution for training GCNs with edge differential privacy. GCON leverages the classic idea of perturbing the objective function to satisfy DP and maintains an unaltered graph convolution process. Our rigorous theoretical analysis offers tight, closed-form bounds on the sensitivity of the graph convolution results and quantifies the impact of an edge modification on the trained model parameters. Extensive experiments using multiple benchmark datasets across diverse settings demonstrate the consistent superiority of GCON over existing solutions.

cs.CR↗

arXiv2023

PrivLava: Synthesizing Relational Data with Foreign Keys under Differential Privacy

Answering database queries while preserving privacy is an important problem that has attracted considerable research attention in recent years. A canonical approach to this problem is to use synthetic data. That is, we replace the input database R with a synthetic database R* that preserves the characteristics of R, and use R* to answer queries. Existing solutions for relational data synthesis, however, either fail to provide strong privacy protection, or assume that R contains a single relation. In addition, it is challenging to extend the existing single-relation solutions to the case of multiple relations, because they are unable to model the complex correlations induced by the foreign keys. Therefore, multi-relational data synthesis with strong privacy guarantees is an open problem. In this paper, we address the above open problem by proposing PrivLava, the first solution for synthesizing relational data with foreign keys under differential privacy, a rigorous privacy framework widely adopted in both academia and industry. The key idea of PrivLava is to model the data distribution in R using graphical models, with latent variables included to capture the inter-relational correlations caused by foreign keys. We show that PrivLava supports arbitrary foreign key references that form a directed acyclic graph, and is able to tackle the common case when R contains a mixture of public and private relations. Extensive experiments on census data sets and the TPC-H benchmark demonstrate that PrivLava significantly outperforms its competitors in terms of the accuracy of aggregate queries processed on the synthetic data.

cs.DB↗