This study presents a novel approach to enhance Large Language Models’ (LLMs) capabilities in cybersecurity analysis through two main methods. First, we propose a Structure-guided Enhancement Network that leverages Finite State Machine (FSM) representations to guide the attention mechanism in processing API call sequences for malware classification. The guided-attention mechanism integrates structural FSM features to enhance the attention computation on API sequences. Second, we develop a Multi-modal Fusion System that aligns assembly code and API call sequences from the same malware through cross-attention mechanisms, creating a shared latent space for improved context generation. Our research demonstrates how structural guidance and multi-modal fusion can enhance LLMs’ performance in cybersecurity-specific tasks.

This paper focuses on patent analysis by using deep learning and customized language model to help human beings to deal with labor-intensive, time-consuming. and knowledge-intensive paperwork. The idea is to train a customized language model. Pat-DistilRoBERTa, who is a patent specialist, and it can map a patent document to a proper high-dimensional vector for later mathematical analysis, therefore, we do not need to deal with text-based data afterward. We anticipate can better understand and represent a patent document. As a comparison, we employ a pre-trained DistilRoBERTa model as a generalist for patent analysis. We conducted numerous experiments on semiconductor patents of leading global companies, specifically TSMC and Samsung. These experiments show that we can easily leverage the language models to appropriately embed a professional technique document for different downstream tasks. For example, Pat-DistilRoBERTa improved the F1-score by 5% to 15% when classifying context and achieved 99% accuracy in classifying patents by IPC. We anticipate that such AI-assisted analysis with specialist language model can assist patent attorneys in reducing their workload as technology improves.

Language models like BERT have advanced the representation of textual semantics in high-dimensional latent spaces, enabling numerous natural language processing applications. However, their capacity to represent domain-specific concepts, such as "color," remains underexplored. This study investigated the semantic representation of text in color concept subspace of latent space. Using embeddings of nearly 1,000 color names from the XKCD color survey generated by BERT, we identified limitations in BERT’s ability to cluster perceptually similar colors. To address this, we proposed a supervised learning approach to project embeddings into a color-specific subspace, isolating and enhancing color semantics. Experimental results demonstrated the methodology’s effectiveness in improving semantic clustering through qualitative and quantitative evaluations. Moreover, our general approach not only explored the concept of color but also provided the possibility of exploring and disentangling semantic subspaces for other domain-specific concepts, contributing to the understanding and manipulation of latent space structures in language models.

Recently, deception detection on human videos is an eye-catching techniques and can serve lots applications. AI model in this domain demonstrates the high accuracy, but AI tends to be a non-interpretable black box. We introduce an attention-aware neural network addressing challenges inherent in video data and deception dynamics. This model, through its continuous assessment of visual, audio, and text features, pinpoints deceptive cues. We employ a multimodal fusion strategy that enhances accuracy; our approach yields a 92% accuracy rate on a real-life trial dataset. Most important of all, the model indicates the attention focus in the videos, providing valuable insights on deception cues. Hence, our method adeptly detects deceit and elucidates the underlying process. We further enriched our study with an experiment involving students answering questions either truthfully or deceitfully, resulting in a new dataset of 309 video clips, named ATSFace. Using this, we also introduced a calibration method, which is inspired by Low-Rank Adaptation (LoRA), to refine individual-based deception detection accuracy.

Nowadays, various video data are springing up. In the field of human-centric video analysis, deception detection becomes a crucial issue to us. Thanks to AI development, automated deception detection has been researched for a while. Nonetheless, the mainstream techniques work as a black box, which is unexplainable. This paper presents an attention mechanism on visual and audio features, which makes the detection interpretable. Besides, we embrace the approach of multi-modal by combining visual, audio, and transcription features as an ensemble model, which can achieve 96% of accuracy.

Not only the number of deployed IoT devices increases but also that of IoT malware increases. We eager to understand the threat made by IoT malware but we lack tools to observe, analyze and detect them. We design and implement an automatic, virtual machine-based profiling system to collect valuable IoT malware behavior, such as API call invocation, system call execution, etc. In addition to conventional profiling methods (e.g., strace and packet capture), the proposed profiling system adapts virtual machine introspection based API hooking technique to intercept API call invocation by malware, so that our introspection would not be detected by IoT malware. We then propose a method to convert the multiple sequential data (API calls) to a family behavior graph for further analysis.

This study explores an important issue of data mining - missing value - by using Neural Network for missing value imputation. We focus on social network data for P2P lending risk assessment. However, social network data usually contain lots of missing values so that predicting accurate credit level is not feasible. We propose a neural network based approach and adapt Generative Adversarial Network concept to generate the missing values. The Generator could output the possible values of the missing values; while the Discriminator could judge if the output values from Generator is appropriate or not for further training. We use real-world P2P data without missing values and the construct datasets with random missing values for evaluation.

Tickets for popular events is usually sold out instantly after they are available online. In addition to lots of demand from the fans, the scalpers try to acquire a large amount of tickets and resell them to earn quick money. At the moment the tickets are available, the web server receives requests from the legitimate buyers and the scalpers. It suffers from a DDoS-like requests and may fail to function. To mitigate this problem, we introduce a Proof-of-Work mechanism with game theory rule on the server that issues a puzzle with different difficulties to a buyer to consume its computation resource. The general buyers and the scalpers both behave according to the PoW puzzle and game theory rules, so that the expected time to access the server will be distributed to different time segments. Thus, the congestion on the web server is mitigated.