Dhananjai Sharma

Ryan Wong, Hosea David Yu Fei Ng, Dhananjai Sharma et al.

Large Language Models (LLMs) remain susceptible to jailbreak exploits that bypass safety filters and induce harmful or unethical behavior. This work presents a systematic taxonomy of existing jailbreak defenses across prompt-level, model-level, and training-time interventions, followed by three proposed defense strategies. First, a Prompt-Level Defense Framework detects and neutralizes adversarial inputs through sanitization, paraphrasing, and adaptive system guarding. Second, a Logit-Based Steering Defense reinforces refusal behavior through inference-time vector steering in safety-sensitive layers. Third, a Domain-Specific Agent Defense employs the MetaGPT framework to enforce structured, role-based collaboration and domain adherence. Experiments on benchmark datasets show substantial reductions in attack success rate, achieving full mitigation under the agent-based defense. Overall, this study highlights how jailbreaks pose a significant security threat to LLMs and identifies key intervention points for prevention, while noting that defense strategies often involve trade-offs between safety, performance, and scalability. Code is available at: https://github.com/Kuro0911/CS5446-Project

Venice Erin Liong, Thi Ngoc Tho Nguyen, Sergi Widjaja et al.

In this paper, we present an Assertion-based Multi-View Fusion network (AMVNet) for LiDAR semantic segmentation which aggregates the semantic features of individual projection-based networks using late fusion. Given class scores from different projection-based networks, we perform assertion-guided point sampling on score disagreements and pass a set of point-level features for each sampled point to a simple point head which refines the predictions. This modular-and-hierarchical late fusion approach provides the flexibility of having two independent networks with a minor overhead from a light-weight network. Such approaches are desirable for robotic systems, e.g. autonomous vehicles, for which the computational and memory resources are often limited. Extensive experiments show that AMVNet achieves state-of-the-art results in both the SemanticKITTI and nuScenes benchmark datasets and that our approach outperforms the baseline method of combining the class scores of the projection-based networks.