Research

This academic paper discusses extending SBOMs (software bill of materials, which are detailed lists of all components and dependencies in software) to create AIBOMs that can describe agentic AI systems (AI systems that can take independent actions and make decisions). The paper proposes schema extensions, methods for coordinating multiple AI agents, and ways to evaluate whether AI systems produce consistent and reproducible results.

Deep neural networks used for image classification are vulnerable to adversarial examples (slightly altered images designed to fool AI models). This paper proposes a new adversarial training method that improves robustness by reducing inter-class feature overlap (the problem where the AI's internal representations of different image categories get too close together), making it harder for adversarial attacks to work. The method was tested on standard image datasets and showed competitive performance compared to existing adversarial training techniques.

This article argues that there are fundamental mathematical limits to how secure and well-aligned (following intended behavior) AI systems can be, and that understanding these limits is important before deploying AI widely. The research also shows that AI systems have basic reasoning limitations that stem from these same information-theoretic constraints.

DreamFuse is a new AI system designed to seamlessly blend foreground objects (like people or animals) into background scenes to create realistic composite images. The system uses a diffusion model (a type of AI that generates images by gradually refining random noise) with a Diffusion Transformer architecture and special attention mechanisms to understand how foreground and background elements should interact, supporting tasks like object insertion, replacement, and editing based on text descriptions.

DrawMotion is a new AI system that generates 3D human motions by combining text descriptions with hand-drawn sketches, addressing the problem that people struggle to describe complex movements using words alone. The system uses a diffusion-based framework (a type of AI model that generates outputs by gradually refining random noise) with three key features: automatic conversion of hand-drawn stick figures into motion, a Multi-Condition Module that efficiently combines text and drawing inputs, and training-free guidance that adjusts the output to match user intentions. User studies show this drawing-based approach cuts the time needed to create desired motions by about 46.7% compared to text-only methods.

This paper addresses transferable targeted attacks (TTAs), which are adversarial techniques that trick AI models by creating modified images that fool multiple different models, not just one. The researchers found that simple scaling transformations (resizing images at different sizes) are surprisingly effective for these attacks and proposed S4ST, a method that combines scaling with other transformations to achieve better results without needing access to the victim model's training data.

This paper presents a new mathematical method for analyzing ragged tensors, which are multidimensional data structures with irregular patterns (unlike regular, grid-like tensors that existing methods assume). The researchers developed a CANDECOMP/PARAFAC (CP)-based approach, a type of tensor decomposition (breaking down complex multidimensional data into simpler factors), that uses a binary weighting system to handle the irregular parts and split the problem into independent, easier-to-solve subproblems. They tested their method on image and biological data and found it outperformed existing approaches in both accuracy and speed.

This research addresses a problem in time-series domain adaptation, where AI models need to work with new data (target domain) without accessing the original training data (source domain) due to privacy concerns. The authors propose TemSR (Temporal Source Recovery), a framework that recreates source-like data patterns to help models transfer their learned temporal dependencies (patterns that change over time) to new datasets, without needing the original source data or special preparation from data owners.

This research paper describes a method for automatically generating password mangling rules (transformations that modify passwords systematically) using adaptive density clustering (a technique that groups similar data points together based on how densely packed they are). The approach aims to improve password security by learning patterns from real password data to create more effective rules for testing password strength.

Researchers developed a transformer-based regression model (a type of deep learning architecture that processes sequential data) that combines movement data, shoe sole images, shoe information, and body measurements to predict how slip-resistant different shoes are on icy surfaces. Testing on 84 different shoes across 578 walking trials showed the model worked better on dry ice than wet ice, suggesting that combining biomechanical data with shoe characteristics could help prevent slip-related injuries.

This paper addresses a problem in cross-corpus speech emotion recognition (SER), which is the task of detecting emotions in speech across different datasets that may have different recording conditions or speaker characteristics. Existing AI methods that generate artificial training samples often create low-quality or noisy data, which weakens the learning process. The authors propose a new model called DFMDM that improves sample quality through three stages: generating samples with target characteristics, filtering out poor-quality samples, and regenerating better samples by mixing good ones with real data.

This research paper presents VM-DSSE-FB, a new encryption method for cloud storage that lets multiple users search and update encrypted data while maintaining privacy. The system uses techniques like symmetric encryption (encoding data with a shared secret key), homomorphic addition (performing calculations on encrypted data without decrypting it first), and bitmap indexing (a method for organizing data searches) to protect against certain attacks and verify that search results are accurate and complete.

MaliVD is a deep learning method that detects vulnerabilities (security flaws) in source code and identifies exactly which lines contain them, using a multi-modal attention mechanism (a technique that lets the AI focus on important parts of code by analyzing it in multiple ways, like looking at the code's sequence, tree structure, and relationships between components). Traditional security tools create too many false alarms and struggle with complex modern software, but MaliVD performs better than eight other detection methods by extracting different types of code features and prioritizing suspicious sections.

Researchers discovered that the SAE J1772 charging control protocol, which manages communication between electric vehicle chargers and cars, lacks proper authentication (verification that devices are who they claim to be). They created ChargeX, a hardware attack that modifies charging control signals to disrupt charging schedules, cause denial of service (DoS, making systems unavailable), or damage batteries. Tests on Tesla chargers and home chargers showed the attacks can force unwanted charging states or crash chargers into error states.

Current privacy regulations fail to prevent harms in data-driven economies because they rely on outdated frameworks and respond too slowly to violations. The paper proposes a new regulatory system based on contextual integrity (the idea that privacy expectations vary depending on the situation and relationships involved) that uses three feedback loops: creating appropriate rules, monitoring compliance in real-time, and continuously assessing whether the system works.

This is a survey paper published in an academic journal that reviews recent progress in conversational data generation, which refers to techniques for creating dialogue datasets (collections of conversations) used to train and improve AI systems. The paper appears to be a comprehensive overview of advances in this field as of July 2026, but no specific technical findings, vulnerabilities, or security issues are described in the provided content.

Researchers have developed AISM (adversarial image steganography model, a technique that hides data inside images while making them resistant to AI recognition), a method for protecting images from being recognized by unauthorized AI systems. The approach uses adversarial techniques (methods that deliberately trick AI models by adding subtle, invisible changes to data) combined with steganography (the practice of hiding information within other data) to prevent unwanted AI analysis while keeping the images visually normal to humans. This work addresses privacy concerns where people want to prevent their images from being processed by AI systems without permission.

This research evaluates 28 large language models on named entity recognition (NER, the task of identifying and labeling people, places, and organizations in text) across 13 datasets to understand how well they perform. The study found that all models experience hallucinations (where the AI generates false or unsupported information), but a two-phase framework called LLM-NER that includes a "Check phase" to verify recognized entities can help reduce these errors.

CShard is a blockchain sharding (dividing transaction processing across multiple parallel groups of nodes) protocol that uses repairable fountain codes (a data encoding method that allows recovery of lost information) to prevent system failures when shards become corrupted. The protocol introduces a ghost reporter mechanism that lets all nodes verify transactions, allowing smaller shard sizes while maintaining security and enabling recovery of corrupted blocks through decoding.