arXiv subjects

Siew Kei Lam

Publications and source records attributed to Siew Kei Lam.

6 recordsLinked to original sources

arXiv2026

Subgroup Membership Inference Audits of Differentially Private Synthetic Text

Synthetic data releases are increasingly proposed in the literature as a means of sharing realistic data replicas in lieu of sensitive private datasets. Even when the worst-case privacy leakage of such releases is bounded by means of differential privacy (DP), in practice a residual risk remains. Membership inference attack (MIA) audits are conducted to empirically quantify this risk. However, existing methods only measure average-case risk for randomly drawn records, which might conceal the risk to vulnerable subgroups. To highlight this issue, we define a subgroup-targeted membership inference game in which the target pool is an explicit parameter, and instantiate it with an audit of 32 proxies under three scenarios with different levels of attacker knowledge, across four datasets, three generators (DP-SGD fine-tuning, API-based prompting, and activation steering), and five privacy budgets. The audit shows that synthetic releases leak subgroup membership and that prior attacks systematically underestimate this leakage. DP is effective at the aggregate level: it substantially reduces average leakage at every budget we test. Three observations temper this picture. First, the remaining leakage is concentrated rather than spread out: under DP, a tenth of the records carries roughly 40% of it. Second, the protection DP delivers in practice is uneven: within its worst-case guarantee, the noise removes more of the measured leakage from random records than from high-risk ones---and a merged-pool audit that scores both record types against shared negatives confirms this at the record level. Third, \emph{which} records leak proves to be a property of the release mechanism rather than of the record alone, so record-level risk cannot be assessed independently of the release.

cs.CR↗

arXiv2026

SynBench: A Benchmark for Differentially Private Text Generation

Synthetic text generation with Differential Privacy (DP) guarantees emerges as a principled approach that can enable the sharing of sensitive datasets across institutional and regulatory boundaries, while bounding the risks of re-identification and membership inference. LLM-based methods deliver promising results; however, comparisons are exacerbated by differing evaluation setups and "private" datasets, potential pre-training contamination is not considered and guarantees are not verified with DP audits. To advance this field, we introduce a unified evaluation framework with standardised utility and fidelity metrics and privacy audits, encompassing nine curated datasets that capture domain-specific complexities such as technical jargon, long-context dependencies, and specialised document structures. In a large-scale empirical study, we benchmark LLM-based state-of-the-art DP text generators of varying sizes (between 1--8B). Our results indicate that DP synthetic text generation remains an unsolved challenge, with quality deteriorating more as the private datasets deviate further from the generators' pre-training corpora. Our novel synthetic text membership inference attack (MIA) explains this observation: Synthetic data quality is overestimated when LLMs have been pre-trained -- without DP -- on portions of the "private" data to be generated. Finally, our work provides the first quantitative evidence that this "public pre-training and private generation" paradigm invalidates the guaranteed privacy bounds of real-world private datasets.

cs.AI↗

arXiv2025

T2V-OptJail: Discrete Prompt Optimization for Text-to-Video Jailbreak Attacks

In recent years, fueled by the rapid advancement of diffusion models, text-to-video (T2V) generation models have achieved remarkable progress, with notable examples including Pika, Luma, Kling, and Open-Sora. Although these models exhibit impressive generative capabilities, they also expose significant security risks due to their vulnerability to jailbreak attacks, where the models are manipulated to produce unsafe content such as pornography, violence, or discrimination. Existing works such as T2VSafetyBench provide preliminary benchmarks for safety evaluation, but lack systematic methods for thoroughly exploring model vulnerabilities. To address this gap, we are the first to formalize the T2V jailbreak attack as a discrete optimization problem and propose a joint objective-based optimization framework, called T2V-OptJail. This framework consists of two key optimization goals: bypassing the built-in safety filtering mechanisms to increase the attack success rate, preserving semantic consistency between the adversarial prompt and the unsafe input prompt, as well as between the generated video and the unsafe input prompt, to enhance content controllability. In addition, we introduce an iterative optimization strategy guided by prompt variants, where multiple semantically equivalent candidates are generated in each round, and their scores are aggregated to robustly guide the search toward optimal adversarial prompts. We conduct large-scale experiments on several T2V models, covering both open-source models and real commercial closed-source models. The experimental results show that the proposed method improves 11.4% and 10.0% over the existing state-of-the-art method in terms of attack success rate assessed by GPT-4, attack success rate assessed by human accessors, respectively, verifying the significant advantages of the method in terms of attack effectiveness and content control.

cs.CV↗

arXiv2022

Reinforced Continual Learning for Graphs

Graph Neural Networks (GNNs) have become the backbone for a myriad of tasks pertaining to graphs and similar topological data structures. While many works have been established in domains related to node and graph classification/regression tasks, they mostly deal with a single task. Continual learning on graphs is largely unexplored and existing graph continual learning approaches are limited to the task-incremental learning scenarios. This paper proposes a graph continual learning strategy that combines the architecture-based and memory-based approaches. The structural learning strategy is driven by reinforcement learning, where a controller network is trained in such a way to determine an optimal number of nodes to be added/pruned from the base network when new tasks are observed, thus assuring sufficient network capacities. The parameter learning strategy is underpinned by the concept of Dark Experience replay method to cope with the catastrophic forgetting problem. Our approach is numerically validated with several graph continual learning benchmark problems in both task-incremental learning and class-incremental learning settings. Compared to recently published works, our approach demonstrates improved performance in both the settings. The implementation code can be found at \url{https://github.com/codexhammer/gcl}.

cs.LG↗

arXiv2020

Graph2Kernel Grid-LSTM: A Multi-Cued Model for Pedestrian Trajectory Prediction by Learning Adaptive Neighborhoods

Pedestrian trajectory prediction is a prominent research track that has advanced towards modelling of crowd social and contextual interactions, with extensive usage of Long Short-Term Memory (LSTM) for temporal representation of walking trajectories. Existing approaches use virtual neighborhoods as a fixed grid for pooling social states of pedestrians with tuning process that controls how social interactions are being captured. This entails performance customization to specific scenes but lowers the generalization capability of the approaches. In our work, we deploy \textit{Grid-LSTM}, a recent extension of LSTM, which operates over multidimensional feature inputs. We present a new perspective to interaction modeling by proposing that pedestrian neighborhoods can become adaptive in design. We use \textit{Grid-LSTM} as an encoder to learn about potential future neighborhoods and their influence on pedestrian motion given the visual and the spatial boundaries. Our model outperforms state-of-the-art approaches that collate resembling features over several publicly-tested surveillance videos. The experiment results clearly illustrate the generalization of our approach across datasets that varies in scene features and crowd dynamics.

cs.CV↗

arXiv2019

Situation-Aware Pedestrian Trajectory Prediction with Spatio-Temporal Attention Model

Pedestrian trajectory prediction is essential for collision avoidance in autonomous driving and robot navigation. However, predicting a pedestrian's trajectory in crowded environments is non-trivial as it is influenced by other pedestrians' motion and static structures that are present in the scene. Such human-human and human-space interactions lead to non-linearities in the trajectories. In this paper, we present a new spatio-temporal graph based Long Short-Term Memory (LSTM) network for predicting pedestrian trajectory in crowded environments, which takes into account the interaction with static (physical objects) and dynamic (other pedestrians) elements in the scene. Our results are based on two widely-used datasets to demonstrate that the proposed method outperforms the state-of-the-art approaches in human trajectory prediction. In particular, our method leads to a reduction in Average Displacement Error (ADE) and Final Displacement Error (FDE) of up to 55% and 61% respectively over state-of-the-art approaches.

cs.CV↗