Research

This research proposes a new certificate management system for UAVs (unmanned aerial vehicles) that uses redactable signatures (digital signatures that can hide specific information) and dynamic accumulators (data structures that track which items are valid) to allow fine-grained revocation at the individual attribute level, while also using a distributed key-generation mechanism (spreading the key creation across multiple authorities rather than relying on one central authority) to reduce the burden of cross-regional certificate management.

PSDO is a privacy-preserving framework for managing energy systems where prosumers (people who both consume and produce energy) can trade power without relying on a central authority. It combines decentralized optimization (a method where multiple parties solve problems together without one central controller) with differential privacy (a mathematical technique that adds noise to data to protect individual information), allowing prosumers to manage their energy autonomously while keeping their data private. Tests on an IEEE 33-bus system showed PSDO can find optimal solutions while protecting privacy better than existing methods.

When multiple humans label the same data for training AI models, they often disagree, creating noise (errors) in the labels that can hurt model performance. This paper shows how to measure how much annotators agree with each other, use that agreement to estimate what kind of label noise exists, and then train models that work better despite the noisy labels. The researchers also provide theoretical guarantees about how well these models will perform on new data.

OwnerHunter is a system that uses large language models (AI trained on vast amounts of text) to identify who owns a website by analyzing webpage content across multiple languages. It improves on older methods that struggled when webpages listed many names or were written in non-English languages, using strategies like checking multiple sources on a page and verifying results to accurately determine the true owner.

Living-off-the-land (LOTL) attacks use legitimate tools already built into a system to avoid being detected by security software. The article examines how attackers could use on-device large language models (AI systems running locally on a user's computer rather than in the cloud) as part of these attacks, though it does not detail specific attack methods or provide concrete defenses.

CLIP and similar vision-language models (AI systems trained on paired images and text to understand both) are vulnerable to adversarial examples (carefully crafted image modifications designed to fool AI systems). Researchers proposed two methods, TGA-ZSR and Comp-TGA, that use text-guided attention (the model's focus on image regions based on text descriptions) to make these models more robust, achieving 9.58% and 11.95% improvements in accuracy when tested on adversarial examples.

Researchers developed AdvDiffusion, a method that creates adversarial patches (special sticker patterns) that can fool face recognition systems into misidentifying people, even in real-world physical environments. The technique uses a diffusion model (an AI that learns to remove noise from images) to generate patches that work against black-box models (AI systems the attacker cannot see inside). These adversarial patches are more effective and transferable across different face recognition systems than previous attack methods.

Split learning (SL, a technique that splits AI model training across multiple computers to reduce computational burden) faces efficiency and security problems in edge computing (distributed computing done on devices near data sources) environments, where slow computers can hinder training and malicious actors may sabotage the model. The paper proposes CoDefend, a framework that uses local epoch regulation (dynamically adjusting how many training rounds each computer performs) and time-aware detection (monitoring for suspicious behavior within specific time windows) to improve both training speed and security while protecting privacy.

This research proposes DSW, a dataset-specific watermarking method to detect when text-to-image diffusion models (AI systems that generate images from text descriptions) are illegally fine-tuned (customized for specific tasks) using protected datasets. The method embeds a hidden watermark image representing the dataset owner into training data, then extracts it from images generated by models that used that data, creating a one-to-one link between the watermark and the specific misused dataset to prove intellectual property theft.

This research paper presents a new method for long-term spatio-temporal forecasting (LSTF, predicting values across space and time over extended periods) using improved graph neural networks (GNNs, AI models that process data organized as connected nodes and edges). The authors address limitations in existing multi-GNN approaches by proposing dynamic multigraph structures that better capture relationships between different data dimensions and use attention mechanisms (techniques that help the model focus on the most important information) to improve prediction accuracy for applications like parking availability and air quality monitoring.

StrokePIN is an authentication system that uses keystroke dynamics (the unique way a person types, including timing and pressure patterns) combined with other data types to verify users' PIN (personal identification number) entries on mobile devices. The system uses a few-shot learning technique called Siamese Network (a machine learning approach that learns from very few examples) to work efficiently without needing to retrain constantly, and it includes security analysis showing that keystroke dynamics can provide meaningful protection against guessing attacks.

This research proposes a federated learning (a decentralized machine learning approach where multiple computers train a model together without sharing raw data) approach for classifying hyperspectral images (images that capture many light wavelengths to reveal details about Earth's surface) across multiple satellite sources. The method addresses two key challenges: improving how well each local computer learns by sharing relevant features with others, and handling bias that arises when different computers have uneven amounts of training data.

This paper describes SPARTA, a protocol designed to let people create multiple separate avatars (digital representations of users in virtual spaces) in the metaverse while keeping those avatars unlinkable, meaning no one can connect different avatars to the same real person. The protocol uses mercurial signatures (a cryptographic technique that allows flexible key usage) and zero-knowledge proofs (ways to prove something is true without revealing how you know it) to enable secure authentication and prevent misuse through a reputation system based on time-based hash chains (sequences of data linked by timestamps).

Ano2Rule is a new method that makes unsupervised anomaly detection models (AI systems that find unusual patterns without being trained on examples of what's normal) more understandable to humans by converting them into simple rules. The approach breaks down how normal data is distributed into multiple parts and creates boundary rules that explain when the model flags something as anomalous (abnormal), making it easier for security experts to trust and deploy these systems in high-stakes situations like detecting network intrusions or protecting IoT devices (internet-connected devices).

Deep neural networks can be attacked through backdoors, where attackers secretly poison training data to make the model misclassify certain inputs while appearing normal otherwise. This paper proposes Cert-SSBD, a defense method that uses randomized smoothing (adding random noise to samples) with sample-specific noise levels, optimized per sample using stochastic gradient ascent, combined with a new certification approach to make models more resistant to these attacks.

Gradient leakage attacks (methods that steal private data by analyzing the mathematical updates sent between computers in federated learning, where AI training happens across multiple devices) pose privacy risks in federated learning systems. Researchers discovered that different layers of neural networks (sections that process information at different stages) leak different amounts of private information, so they created Layer-Specific Gradient Protection (LSGP), which applies stronger privacy protection to layers that leak more sensitive data rather than protecting all layers equally.

When users send prompts to LLM services like ChatGPT, sensitive personal information (such as names, addresses, or ID numbers) can leak out, even when basic privacy protections are used. This paper presents Rap-LI, a framework that identifies which parts of a user's input contain sensitive data and applies stronger privacy protection to those specific parts, rather than treating all data equally.

This paper introduces FOFedAvg, an improved version of federated learning (a technique where multiple remote computers train an AI model together while keeping their data private) that uses fractional-order stochastic gradient descent (a math-based method that remembers past training steps to learn better). FOFedAvg addresses common problems in federated learning like slow training speed, heavy communication costs, and issues when different clients have different types of data by using memory-aware updates that improve both efficiency and stability.

Researchers created FAOSPA, an attack method that exploits weaknesses in video anomaly detection (VAD, systems that automatically identify unusual events in surveillance footage) by manipulating the frequency domain (the mathematical representation of signals broken down by their different oscillation rates). The attack severely degrades VAD system performance, reducing detection accuracy from 91% to 34% with minimal changes to video data, and the researchers emphasize that stronger defenses are urgently needed.