32584 research outputs found
Sort by
Enhancing Robustness of CNN Model for Malware Detection using GAN-Based Data Augmentation and Transfer Learning
No full textMalware classification is a critical component in the field of cybersecurity. Accurate identification of a malware family can enable timely threat detection and response. In this thesis, we propose a robust image-based malware classification pipeline using Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs), with a focus on improving performance for underrepresented malware families. We train a baseline CNN model on the Malimg dataset across 25 malware families, but observe misclassifications in classes with limited data and overlapping visual features. To address this, we apply targeted augmentations and generate class-specific synthetic data using StyleGAN2-ADA. A CNN trained on the combination of real and synthetic data outperforms both the baseline CNN and fine-tuned transfer learning models. This project highlights the impact of data augmentation and the importance of preserving malware patterns during preprocessing to enhance malware classification performance
Adaptive Cobot Interaction via Smartwatch Data Fusion for Car Assembly Automation
No full textIn modern car manufacturing, collaborative robots (cobots) work with human operators during shared workcell interactions to maximize production speed and flexibility. Collaboration between humans and robots is safe and effective only when operator intent recognition via a single wrist-worn inertial measurement unit (IMU) is accurate and low-latency. This thesis develops an IMU-only intent recognition pipeline, and is evaluated on three datasets: the public OPPORTUNITY dataset, the Sony Smartwatch Gesture dataset and a custom Samsung Galaxy watch 6 dataset. The proposed framework leverages five step sequence-to-label problems which are stepwise posed as data streams transforming raw IMU data into trainable tensors. For that purpose, a deterministic seven-step transformation consists of transport-agnostic ingestion, gap interpolation, low-pass filtering, and on-the-fly label-uncorrupted znormalisation. Z-normalised epochs of 100 frames at sampling frequency 50 Hz are fed to five sequence-deep architectures. The BiLSTM, GRU, 1D CNN, CNN+LSTM, and a four-block Temporal Convolutional Network TCN were trained and evaluated under stratified 80/20, Leave-One-Subject-Out (LOSO), Leave-One-Experiment-Out (LOEO), and stratified K-fold validation—all separately on each dataset. The TCN consistently achieves the best trade-off between accuracy and efficiency: a macro-F1 score of 0.923 on the OPPORTUNITY 80/20 split, 0.967 under LOSO on the Sony corpus, and 0.427 under LOSO on OPPORTUNITY using only 200k parameters while sustaining sub-20 ms inference on standard CPUs. A Bayesian hyperparameter search on the SJSU HPC cluster demonstrates a 40% reduction in tuning time, showcasing scalable reproducibility across 1,232 CPU cores and multiple GPU nodes
TraceAI: Intelligent Distributed Tracing Using Large Language Models
No full textDistributed systems are difficult to trace using traditional methods due to the scale of data volume and complexity, and they usually require a lot of manual analysis. TraceAI tries to solve these problems by integrating Large Language Models with the tracing tools to automatically enhance the trace data evaluation. The project aims to provide an AI-driven solution for monitoring and understanding the flow of requests across services, anomaly detection, root cause analysis and performance optimization. It can thus automate finding out systems problems using LLMs thereby carrying out large scale trace data analysis. Anticipated results from the effort will be a prototype AI (Artificial Intelligence) Chatbot enhancing efficiency of tracing workflow significantly that performs the given requirements faster, making systems reliable while ensuring higher performance from a complex distributed setup
SemanticGraphRec: Lightweight Hybrid Recommendations Powered by Semantic Item Representations and Graph Collaborative Filtering
No full textGraph neural networks (GNNs) have emerged as a powerful paradigm for collaborative filtering. However, they often fall short in fully leveraging side textual content, resulting in suboptimal recommendations. To address this limitation, we explore the synergy between GNNs and deep contextual embeddings of item descriptions, aiming to enhance recommendation quality on the Amazon-Books dataset. We propose SemanticGraphRec, which combines GNNs with Large Language Models (LLMs) to leverage both collaborative filtering and textual item content. Experimental results demonstrate that incorporating semantic item embeddings produced by fine-tuning LLMs consistently improves performance. Our approach enhances recommendation relevance in sparse data scenarios by leveraging both textual content and graph structure, offering a promising direction for more context-aware and personalized recommender systems across diverse application domains
Machine Learning based Network Traffic Classification with Cosine-Similarity based Out-Of-Distribution Detection
No full textThe changes occurring in the amount of encrypted network traffic is growing at an alarming rate. This development has created intricate problems in traffic classification which is vital for effective cybersecurity. Moreover, most frameworks seem to ignore OOD detection, model calibration and novel pattern detection as cornerstone problem areas. The due analysis is presented as a machine learning approach aimed at resolving encrypted traffic classification issues and focuses on novel OOD detection and calibration issues. Primary contributions comprise detection of out-of-distribution states using softmax scaled cosine similarity, advanced variance-based feature elimination, and lowering ECE using stringent NNs. This work demonstrates that ensemble/models using a blend of these techniques using a dataset of VPN encrypted and unencrypted network traffic achieve reliable and high-accurate performance across different classes of traffic. The results underscore the impact of the proposed solutions on improving dynamics, reliability, and interpretability of machine learning models and through them secure and intelligent monitoring of network traffic
GaLoRA: A Lightweight Graph-Aware LLM Framework for Node Classification on Text-Attributed Graphs
No full textWith the exponential rise of language models (LMs) and their potential to understand semantic relationships, large LMs are being used across a wide range of applications. Text-attributed graphs (TAGs) are one notable example where LLMs can be combined with Graph Neural Networks (GNNs) to enhance node classification results. TAGs associate textual content with each node and are commonly seen in various domains such as social networks, citation graphs, recommendation systems, etc. Effectively modeling TAGs would enable deeper insights into different aspects of the graph and improve decision-making in relevant domains. We present GaLoRA, a parameter-efficient framework to integrate structural information in large LMs. GaLoRA demonstrates a strong performance for the node classification task on TAGs, performing on par with state-of-the-art models while requiring fewer trainable parameters. We experiment with three real-world datasets to showcase GaLoRA’s effectiveness in combining structural and contextual information of TAGs
Two-Stage Emotion Detection from Multimodal Data
No full textEmotion detection plays a crucial role in human-computer interaction, enabling machines to recognize and respond appropriately to human emotional states. This project explores a two-stage approach to emotion detection using multimodal data, first predicting dimensional values (Arousal, Valence, Dominance) from textual and audio inputs, then mapping these representations to discrete emotion categories. We compare this approach with direct categorical classification using transformer-based language models like BERT, RoBERTa, and DeBERTa for text processing, alongside various audio feature extraction methods including MFCCs and spectrograms. Using the IEMOCAP dataset, we evaluate both approaches across text-only, audio-only, and multimodal configurations. Our findings reveal that while the two-stage approach provides richer emotional representations, direct classification achieves superior accuracy (91.82% with RoBERTa) compared to the two-stage method (90.13% with the same model). Interestingly, text-only approaches slightly outperform multimodal ones, though the gap narrows with optimal fusion strategies. For dimensional prediction, we observe that textual features better capture valence (positive/negative sentiment), while audio features more effectively represent arousal (emotional intensity). This research contributes valuable insights into the tradeoffs between dimensional and categorical approaches to emotion recognition, with implications for applications requiring either maximum classification accuracy or nuanced emotional understanding. The findings suggest that application requirements should dictate the choice between these approaches, with direct classification preferred for accuracy-critical tasks and the two-stage approach for scenarios benefiting from continuous emotional representation
EvoRGCN: Harnessing ESM-2 Evolutionary Embeddings with Relational GCNs for High-Fidelity Protein-Protein Interaction Prediction
No full textAccurately predicting protein-protein interactions (PPIs) is essential for understanding cellular function and advancing biomedical discovery. We model PPIs as graphs, where nodes represent proteins and edges denote interactions. Using interaction data from the STRING database, we use two samples of it, namely the benchmark datasets—SH27K and SH148K—filtered by confidence score and annotated by interaction mode (multiple relations). In this project, we present EvoRGCN, a graph-based machine learning framework for PPI prediction that integrates both sequence-level (ESM-2 embeddings) and network-level information. We incorporate various Graph Neural Network architectures, including Graph Convolutional Networks, Graph Attention Networks, and Relational Graph Convolutional Networks. Our experiments systematically evaluate the undirected vs. the directed model, edge semantics (single-mode vs. multi-relational), and various node feature types (one-hot encodings vs. ProtBERT vs. ESM-2 embeddings) on the prediction accuracy. Results show that using ESM-2 embeddings in a directed multi-relational RGCN gives the best results. These findings highlight the effectiveness of combining graph-based learning with transformer-derived protein features and provide a scalable, interpretable framework for computational PPI prediction
Effects of Data Augmentation on Sponge Identification Using Computer Vision Models
No full textCoral reefs can be primarily found in tropical and sub-tropical regions of our oceans, providing a thriving habitat for millions of species. Marine sponges, which can be frequently found in coral reefs, play a critical role that contributes to the maintenance of these ecosystems, including the recycling of nutrients through water filtration. However, rising ocean temperatures and acidification due to climate change have resulted in the bleaching and death of coral reefs worldwide. In order to preserve these reefs and the sponges that depend on them, scientists have been performing studies on their biodiversity. This includes collecting numerous images of autonomous reef monitoring structures (ARMS) in order to document the species living in an area. Although this method can give scientists a good idea of the biodiversity in a region, the sheer volume of images collected makes it difficult to quickly and accurately analyze entire datasets. To address this challenge, we employed a Mask Regional Convolutional Neural Network (Mask R-CNN) framework to create a model for automating the process of identifying specific sponges. After training a model on previously selected samples of a unique sponge species, we are able to reliably use this model to identify additional instances of the sponges in other images. We then tested how data augmentation on our training sets affects the reliability of these models. These tests demonstrate the ability of Mask R-CNNs to identify sponges and further introduce the possibility of applying similar models to other species and images
RIFT - Reddit Information Falsity Tagger
No full textSocial media platforms such as Reddit are widely used for sharing and consuming information. User-generated content poses a great risk for misinformation creation and dissemination on these platforms. “Fake news”, as it is commonly referred to, has far-reaching social implications, swaying public perception, making political viewpoints more radical, and adversely impacting health decisions. The covariable features that come with fake news make it even harder to detect because it is presented in the form of text, images, videos, and even social interactions. This paper describes a novel method for detecting fake news on Reddit: RIFT, short for Reddit Information Falsity Tagger. The proposed system models Reddit using a Neo4j graph database and employs a graph approach for fake news detection. Users, posts, and subreddits are mapped to nodes with edges like POSTED_BY and BELONGS_TO defining relationships. Each post node is augmented with multimodal features like text embeddings (BERT), image embeddings (CLIP), and upvote/comment engagement levels (upvotes, comments). RIFT is tested on two datasets: the original Fakeddit benchmark and FakedditEnhanced, an extended version created for this project. The enhanced dataset is enriched with more metadata and image embeddings obtained through Reddit’s API which allows for better multimodal detection. The results achieved with Graph Neural Networks (GCN, GraphSAGE) demonstrate the power of applying both content and structure in identifying misinformation