Large Language Models (LLMs) have achieved significant performance in various natural language processing tasks but also pose safety and ethical threats, thus requiring red teaming and alignment processes to bolster their safety. To effectively exploit these aligned LLMs, recent studies have introduced jailbreak attacks based on multi-turn dialogues. These attacks aim to prompt LLMs to generate harmful or biased content by guiding them through contextual content. However, the underlying reasons for the effectiveness of multi-turn jailbreaks remain unclear. Existing attacks often focus on optimizing queries and escalating toxicity to construct dialogues, lacking a thorough analysis of the inherent vulnerabilities of LLMs. In this paper, we first conduct an in-depth analysis of the differences between single-turn and multi-turn jailbreaks and find that successful multi-turn jailbreaks can effectively disperse the attention of LLMs on keywords associated with harmful behaviors, especially in historical responses. Based on this, we propose ASJA, a new multi-turn jailbreak approach by shifting the attention of LLMs, specifically by iteratively fabricating the dialogue history through a genetic algorithm to induce LLMs to generate harmful content. Extensive experiments on three LLMs and two datasets show that our approach surpasses existing approaches in jailbreak effectiveness, the stealth of jailbreak prompts, and attack efficiency. Our work emphasizes the importance of enhancing the robustness of LLMs' attention mechanism in multi-turn dialogue scenarios for a better defense strategy.

With the increasing availability of trajectory data, it is important to have good indexes to facilitate query processing. In this work, we propose BT-Tree, which is built through a recursive bi-partitioning approach, for the processing of range and KNN queries for past trajectory data. We first propose a cost function based method (CFBM) to build the BT-Tree. Specifically, we design a novel cost function, which incorporates the characteristics of both the data and historical query workload, to decide how to partition a BT-Tree node. Then we propose a reinforcement learning (RL) based method to address CFBM's limitations, such as making locally optimal decisions that may lead to global suboptimality. Experiments on three real datasets with up to 800 million data points show that the CFBM generally outperforms the baselines in terms of query processing time and the RL based method consistently outperforms the baselines and has more significant advantages on larger datasets.

Learned indexes have been proposed to replace classic index structures like B-Tree with machine learning (ML) models. They require to replace both the indexes and query processing algorithms currently deployed by the databases, and such a radical departure is likely to encounter challenges and obstacles. In contrast, we propose a fundamentally different way of using ML techniques to build a better R-Tree without the need to change the structure or query processing algorithms of traditional R-Tree. Specifically, we develop reinforcement learning (RL) based models to decide how to choose a subtree for insertion and how to split a node when building and updating an R-Tree, instead of relying on hand-crafted heuristic rules currently used by the R-Tree and its variants. Experiments on real and synthetic datasets with up to more than 100 million spatial objects show that our RL based index outperforms the R-Tree and its variants in terms of query processing time.