Search for: All records

Recent advances in Large Language Models (LLMs) have demonstrated significant potential in the field of Recommendation Systems (RSs). Most existing studies have focused on converting user behavior logs into textual prompts and leveraging techniques such as prompt tuning to enable LLMs for recommendation tasks. Meanwhile, research interest has recently grown in multimodal recommendation systems that integrate data from images, text, and other sources using modality fusion techniques. This introduces new challenges to the existing LLM-based recommendation paradigm which relies solely on text modality information. Moreover, although Multimodal Large Language Models (MLLMs) capable of processing multi-modal inputs have emerged, how to equip MLLMs with multi-modal recommendation capabilities remains largely unexplored. To this end, in this paper, we propose the Multimodal Large Language Model-enhanced Sequential Multimodal Recommendation (MLLM-MSR) model. To capture the dynamic user preference, we design a two-stage user preference summarization method. Specifically, we first utilize an MLLM-based item-summarizer to extract image feature given an item and convert the image into text. Then, we employ a recurrent user preference summarization generation paradigm to capture the dynamic changes in user preferences based on an LLM-based user-summarizer. Finally, to enable the MLLM for multi-modal recommendation task, we propose to fine-tune a MLLM-based recommender using Supervised Fine-Tuning (SFT) techniques. Extensive evaluations across various datasets validate the effectiveness of MLLM-MSR, showcasing its superior ability to capture and adapt to the evolving dynamics of user preferences. 

Scripts on webpages could steal sensitive user data. Much work has been done, both in modeling and implementation, to enforce information flow control (IFC) of webpages to mitigate such attacks. It is common to model scripts running in an IFC mechanism as a reactive program. However, this model does not account for dynamic script behavior such as user action simulation, new DOM element generation, or new event handler registration, which could leak information. In this paper, we investigate how to secure sensitive user information, while maintaining the flexibility of declassification, even in the presence of active attackers-those who can perform the aforementioned actions. Our approach extends prior work on secure-multi-execution with stateful declassification by treating script-generated content specially to ensure that declassification policies cannot be manipulated by them. We use a knowledge-based progress-insensitive definition of security and prove that our enforcement mechanism is sound. We further prove that our enforcement mechanism is precise and has robust declassification (i.e. active attackers cannot learn more than their passive counterpart).