Research

This research paper addresses how to create adversarial examples (inputs slightly modified to fool AI models) that work against unknown target models by testing them on surrogate models (known substitute models first). The authors provide a theoretical framework showing that adversarial examples are more likely to transfer across models when they optimize toward flat minima (solutions that are robust to small changes) while accounting for differences between surrogate and target models, and they propose a new algorithm called DRAP that improves this transferability by creating diverse surrogate models.

This is an academic survey paper that reviews different prompting frameworks, which are structured approaches to asking large language models (AI systems trained on huge amounts of text) questions or giving them instructions to complete tasks. The paper, published in a major computer science journal, catalogues and analyzes various methods researchers have developed to improve how effectively people interact with and get useful results from LLMs.

This research paper describes a new AI method for identifying rock types in oil and gas wells using well log data (measurements from instruments that observe underground rocks). The method addresses the challenge of having very few labeled examples by using adversarial learning (a technique where two AI models compete to improve each other) to reduce differences both within individual wells and across different wells, along with a dynamic labeling mechanism (a system that automatically adjusts which training examples are most reliable) to improve accuracy.

Chat-Scene++ is a framework that helps multi-modal large language models (MLLMs, AI systems that understand both text and images) better understand 3D scenes by organizing them as sequences of objects with contextual information. Unlike previous methods that struggled with identifying specific objects and understanding how they relate to each other in 3D environments, Chat-Scene++ uses pre-trained encoders to extract rich features about objects and their relationships, enabling better performance on tasks like object identification, scene description, and spatial reasoning without needing task-specific customization.

Version 5.5.0 adds new security techniques documenting threats to AI systems, including AI agent tool poisoning (when attackers corrupt tools that AI agents use), supply chain attacks, and cost harvesting (depleting computing resources through expensive queries). It also updates existing techniques and mitigations related to code signing and monitoring AI agent behavior.

This research paper examines how smartphone users develop privacy concerns about location tracking through a 'triple calculus model' (a framework showing how people weigh risks and benefits of sharing location data). By studying 559 smartphone users, researchers found that users' sense of control over location sharing significantly influenced how they perceived both the risks and benefits of location disclosure, and that social influences and past experiences with privacy breaches also shaped their privacy concerns.

This research studies how making AI chatbots seem more human-like (anthropomorphism) affects whether people actually share personal information with them. The study found that while human-like design can build trust and reduce worry about privacy, it can also create an "uncanny valley" effect (where something looks almost human but feels unsettling), and people's actual sharing behavior doesn't always follow what they say they intend to do.

This research proposes a framework that combines reinforcement learning (RL, a machine learning technique where a system learns by trial and error and receiving rewards), control barrier functions (CBFs, mathematical tools that keep a system operating within safe limits), and active fault-tolerant control (AFTC, a method that detects and responds to component failures in real time) to help multi-agent systems (groups of coordinated AI agents or robots) track targets accurately even when their actuators (motors or mechanisms that create movement) fail. The approach ensures safety by maintaining a predefined safe operating zone while using a computationally efficient method suitable for large-scale deployments.

This research addresses a security challenge where malware (malicious software) uses polymorphic evasion tactics, splitting its behavior across multiple processes (concurrent program instances) to evade detection systems. The authors propose a defense framework that uses entropy-based partitioning (dividing system activity logs by information density) and model fusion (combining multiple AI models trained differently) to better detect such multi-process malware, showing improved detection accuracy even when malware attempts advanced evasion techniques.

Text-to-image models (AI systems that generate pictures from written descriptions) can be misused to create unsafe content like sexually explicit or violent images. PromptGuard is a new safety technique that uses a soft prompt (a special text input optimized for safety that works within the model's internal text processing layer) to moderate unsafe requests and prevent the generation of such content while still producing high-quality normal images.

This research addresses a problem where adversarial training (a method to make AI models resistant to adversarial attacks, which are carefully crafted inputs designed to fool the model) works poorly when training data is imbalanced, meaning some classes have many examples while others have very few. The authors propose Tail-Aware Dynamic Adversarial Training (TAD-AT), which improves robustness by adjusting the training loss, attack strategy, and weight averaging to account for which classes are most vulnerable to attacks, rather than just how many examples exist per class.

This research proposes new methods for fine-tuning (customizing a trained AI model for specific tasks) large language models while protecting sensitive data using differential privacy (a technique that adds noise to data to prevent identifying individuals). The paper introduces DP-ZOSO and DP-ZOPO, which use zeroth-order gradient approximation (estimating how to improve the model without calculating exact mathematical directions) instead of traditional methods, making the process faster and more scalable while maintaining privacy protection.

RanDS is a new large-scale dataset containing raw binary files (the compiled machine code of programs) and extracted features designed to help researchers study and detect ransomware (malicious software that encrypts victims' files and demands payment). This resource aims to support the development and testing of machine learning models that can identify ransomware threats more effectively.

Researchers discovered a new backdoor attack (a security flaw where hidden malicious code is planted in training data) on Graph Neural Networks, or GNNs (AI models designed to understand interconnected data). The attack uses a single trigger node (a specially crafted fake data point) attached to a target node to trick the GNN into making wrong predictions not just on that node, but also on its immediate neighbors, while remaining stealthy and achieving over 95% success rates even against existing defenses.

This paper presents a new reinforcement learning (RL, a machine learning technique where an AI learns by receiving rewards for good decisions) method called MoEC for autonomous driving that can handle multiple competing objectives, like safety and efficiency, simultaneously. The authors improved RL for self-driving cars by using multiple evaluation networks (instead of one) to assess different driving goals and by creating a hybrid action space (a mix of abstract guidance and concrete commands) so the car can drive more smoothly and flexibly. Testing showed their method successfully learned to drive well on highway scenarios while balancing efficiency, smooth actions, and safety.

Graph Neural Networks (GNNs, AI systems designed to work with interconnected data structured as graphs) used in graph self-supervised learning (training without labeled data) can be secretly compromised by backdoor attacks (where hidden malicious instructions are embedded in the model). Researchers developed GDetox, a defense method that removes these backdoor features from compromised encoders (the parts of the model that learn to represent data) using knowledge distillation (a technique where a teacher model teaches a student model to learn better), reducing successful attacks to 4% while keeping the model's normal performance nearly unchanged.

Deepfake technology can create fake facial images that are hard to distinguish from real ones, posing risks to privacy and security. This paper proposes a new detection method using Visual Language Models (VLMs, AI systems that understand both images and text) combined with component-specific prompt tuning (customizing input instructions to focus on specific facial parts like eyes and nose). The approach transforms deepfake detection into a Visual Question Answering task and uses a Q-Former module (a feature extraction component guided by instructions) to help the model identify forgery traces in local facial features, achieving better accuracy than existing methods.

Many software organizations claim to make security a priority, but they don't actually provide developers with the tools, training, or culture needed to build secure code. A global survey found significant gaps between what companies say about security and what they actually do to support developers in writing secure software.

This research paper addresses security and transparency challenges in cloud storage for UAV (unmanned aerial vehicle) data by proposing PATD, a system that combines privacy-preserving auditing with transparent deduplication. The paper identifies two main problems: verifying that outsourced data hasn't been corrupted or tampered with (without revealing the data itself), and ensuring that file deduplication (removing duplicate copies to save storage) is performed honestly and transparently by the cloud provider.