Research

TWReID is a system that identifies people using radar signals instead of cameras, even when they're walking freely behind walls or obstacles. It uses MIMO radar (a radar system with multiple transmitters and receivers that send signals in different directions) to detect human movement patterns and match them to identify individuals, achieving high accuracy rates in tests with up to 14 people.

Researchers developed WPEBA, a new type of adversarial attack (a method to trick visual recognition systems by adding subtle noise to images) that uses wavelet packet decomposition (breaking images into different frequency patterns) to fool AI models with very few queries (attempts to test the system). The attack combines multiple surrogate models (practice AI systems used to design the attack) and achieves a 99% success rate while remaining effective even against defended models and commercial APIs (pre-built services).

This research paper presents new methods for creating differentially private CDFs (cumulative distribution functions, which describe how data is distributed), using techniques like polynomial projection and sparse approximation. The approach protects individual data privacy while still allowing accurate statistical analysis, and works well with streaming data and multiple variables.

Researchers created a method called 3DGAA that generates adversarial wraps (deceptive visual coverings) for vehicles to test vulnerabilities in camera-based perception systems used by autonomous cars. The technique uses 3D Gaussian splatting (a method for representing 3D scenes that maintains visual consistency from different angles) to design wraps that fool object detection systems while remaining physically realistic and printable, testing autonomous vehicle safety across different lighting and viewing angles.

This research describes PB-Net v2, an improved AI system that authenticates users by analyzing their hand gestures, which contain unique physiological (body structure) and behavioral (movement patterns) characteristics. The system uses neural networks (computer models inspired by how brains work) to extract and combine these characteristics more effectively than previous versions, achieving better accuracy in recognizing legitimate users and resisting fake gesture attacks.

AuditML is a framework that lets multiple parties run machine learning inference while keeping their data private and allowing anyone to verify the results are correct. It uses arithmetic secret sharing (a technique where data is split into random pieces distributed across parties) and separates the computation phase from the verification phase, so checking the results doesn't slow down the actual inference and can happen publicly afterward.

Cloud-based LLM services usually require users to send their text inputs unencrypted, which creates privacy risks. This paper presents PI-SAFE, a framework that protects privacy during LLM inference (the process of running a model to generate outputs) by splitting the model between client and server and sending obfuscated intermediate representations (partially hidden data states) instead of raw text, combined with adversarial fine-tuning (training the model with specially crafted examples to make it harder to attack) to prevent attackers from reconstructing the original input.

This paper addresses Federated Generalized Category Discovery (Fed-GCD), a task where multiple clients collaborate privately to identify both known and unknown categories in unlabeled data while keeping their data private. The authors propose PCGL (Personalized Contrastive Graph Learning), a framework that separates generic knowledge (shared across all clients) from personalized knowledge (specific to each client) to improve performance for both individual clients and the overall shared model, solving a problem where existing methods create weaker local models by forcing uniform aggregation (combining all client models together) that causes knowledge conflicts.

Researchers developed GECOMP, a method that uses reinforcement learning (a technique where an AI learns by receiving rewards for good actions) to generate adversarial examples (inputs designed to trick AI models) against natural language processing systems. The method creates perturbations (small changes to text) using a library of possible edits and an LLM (large language model) generator, balancing the goal of fooling the target model while maintaining text quality and minimizing the number of queries needed to test it.

This research describes a system for detecting whether people are present in rooms using Wi-Fi Channel State Information, or CSI (data about how Wi-Fi signals travel through space). The system addresses real-world challenges like interference and unreliable connections by using a reference-based denoising method (filtering out unwanted signal noise by comparing against empty-room patterns) and a multi-region strategy that combines data from multiple Wi-Fi links across different areas to improve accuracy in security monitoring applications.

ATLAS v2026.08 is an updated knowledge base documenting adversary tactics and techniques involving AI systems, including attacks against AI-enabled systems and abuse of AI capabilities, based on real-world observations and security research. The update adds new techniques related to autonomous AI agents (such as reconnaissance, attack coordination, and communication between agents), new mitigations for controlling AI agent behavior, and case studies of actual AI-related attacks on infrastructure and government systems.

Cyber risk has become an important concern for top organizational leaders, as attacks grow more complex and make it harder for companies to stay secure while running their business. Research from over 1,800 simulation exercises shows that decision-making gets significantly worse as threats become more complicated, but four management strategies can help: adapting quickly to changes, learning from each incident, making cyber risk everyone's responsibility across the company, and building security capabilities before problems happen.

This academic paper examines security attacks targeting industrial control systems (ICS, which are computers that manage factories, power plants, and other critical infrastructure) when they communicate over wireless networks. The research, published in December 2026, explores vulnerabilities in these wireless connections that attackers could exploit to disrupt or damage essential infrastructure.

This is a systematic review article that examines security and privacy-preserving mechanisms used in collaborative machine learning (where multiple organizations or parties train AI models together while protecting their sensitive data). The article organizes existing approaches into a novel taxonomy, helping researchers and practitioners understand different methods for keeping data secure during collaborative AI training.

A study of over 1,000 first-year students at Bocconi University found that access to ChatGPT (a large language model, or LLM) improved the quality and professionalism of student work on a business assignment, while separate training in causal reasoning (a form of critical thinking involving understanding cause-and-effect relationships) led students to generate more original and diverse ideas. Students who received both ChatGPT access and critical-thinking training showed benefits from each approach, suggesting that AI tools and thinking skills are complementary rather than competing.

This survey article examines how generative AI (machine learning models that can create new content) is being used to produce deepfakes (synthetic media where a person's face or voice is digitally manipulated to appear authentic) and discusses methods to detect them. The paper reviews current techniques for both creating and identifying deepfakes, and considers future challenges in an era where AI-generated content is becoming increasingly sophisticated and difficult to distinguish from real media.

This academic survey examines how AI and intelligent technologies affect security in IoV (Internet of Vehicles, where connected cars communicate with each other and infrastructure). The paper discusses integrating edge computing (processing data closer to vehicles rather than in distant data centers) with AI to improve IoV security, though specific vulnerabilities and their fixes are not detailed in this overview.

This is a research survey paper published in ACM Computing Surveys that compares security risks across different types of AI systems, including LLMs (large language models, which are AI systems trained on massive amounts of text), AI agents (systems that can take actions based on their decisions), and embodied agents (AI systems that interact with the physical world through robots or similar devices). The paper examines and contrasts the various security vulnerabilities and threats that each of these AI system types faces.

Researchers have developed a method for encrypting images using deep learning combined with chaotic systems (mathematical systems that produce unpredictable, random-looking outputs). The approach reconstructs chaotic patterns using neural networks to create a security system that scrambles images in a way that makes them unreadable without the correct decryption key.