Academic papers, new techniques, benchmarks, and theoretical findings in AI/LLM security.
This research paper analyzes inconsistencies in CVSS scores (numerical ratings that measure how serious software vulnerabilities are) within the NVD (National Vulnerability Database, a public repository of known security flaws). The study found that the same vulnerability often receives different CVSS scores depending on which scoring standard or organization assigns the rating, revealing a fragmentation problem in how vulnerability severity is measured and reported.
PUFZIN is a blockchain-IoT (Internet of Things, the network of connected devices) security system that combines PUFs (physical unclonable functions, unique hardware-based identifiers that are hard to forge) with zero-knowledge proofs (a cryptographic method where one party proves knowledge of something without revealing the actual information) to create a secure and scalable network. The research, published in July 2026, addresses how to protect IoT devices and blockchain systems from unauthorized access and tampering.
This academic paper presents a framework for protecting unmanned ground vehicles (UGVs, which are robots that operate on land without human drivers) against cyber attacks by combining offensive and defensive security strategies. The research, published in Computers & Security, addresses how to both defend UGVs from threats and identify vulnerabilities through coordinated security approaches.
VaultFS is a file system (the software layer that manages how files are stored and organized on a computer) that ensures data integrity (accuracy and trustworthiness of stored information) by implementing write-once storage at the file system level, meaning files can only be written once and cannot be modified afterward. This approach protects against accidental or malicious changes to critical data by making it impossible to overwrite or alter files after they are created.
This research paper presents DWT-AMSA, a new method for image steganography (hiding secret data inside images so others cannot detect it) that uses frequency-domain adaptive masking (adjusting which parts of an image's mathematical representation are modified based on the image content) and progressive adversarial training (a machine learning technique where two competing AI systems improve each other iteratively to make the hidden data harder to detect). The method aims to make hidden information more robust and harder for attackers to discover or remove.
This research paper examines how Security Operations Centers (SOCs, teams that monitor and respond to security threats) can work effectively with AI systems by using adaptive trust mechanisms. The study focuses on building resilient operations, meaning systems that can continue functioning even when problems occur, through better collaboration between human security experts and AI tools that can process large amounts of data quickly.
This research paper validates a model for identifying and managing data security risks when higher education institutions use mobile cloud computing (storing and accessing data through mobile devices and internet-based servers rather than local computers). The study empirically tests this security risk model to help universities better understand and protect sensitive data in mobile cloud environments.
Federated learning (a system where multiple computers train an AI model together while keeping their data private) can be unfair to some participants and vulnerable to attacks where bad actors tamper with the process. FairRoP is a new method that uses adaptive client selection (choosing which computers to include based on their trustworthiness) and a bandit algorithm (a technique for balancing exploration and exploitation in decision-making) to improve both fairness and robustness against attacks. The approach combines three components: fairness awareness, attack detection, and q-Balance to handle the different challenges involved.
This article addresses how radar systems (devices that detect objects using electromagnetic waves) can better allocate their time between searching for targets and tracking known ones when facing jamming (intentional interference meant to disrupt detection). The researchers propose a dynamic scheduling strategy using receding-horizon optimization (a method that repeatedly solves shorter planning problems instead of one big long-term problem) combined with mathematical techniques to keep radar performance strong even under jamming attacks.
This research paper proposes E2E-PP, a system that protects privacy in mobile crowdsensing (collecting data from many mobile devices) by combining compressive sensing (a technique that reduces data size while preserving important information) with personalized differential privacy (a method that adds customized noise to data to prevent identifying individuals). The system aims to let mobile devices share sensor data for collective purposes while keeping personal information private.
This academic paper analyzes how Internet of Things devices (smart devices connected to the internet, like security cameras or smart home systems) receive and install software updates. The research examines the mechanisms these devices use to stay current with security patches and new features. The publication appears in a peer-reviewed security journal and was made available online in May 2026.
This research paper proposes DEGAN, a machine learning approach using dual-enhanced GAN (generative adversarial network, a type of AI that learns by having two competing neural networks) to detect botnets (networks of infected computers controlled remotely) in IIoT (industrial internet of things, devices like sensors and machines in factories connected to the internet). The method addresses the challenge of imbalanced data, where there are far fewer examples of botnet attacks than normal network activity, which makes training detection systems difficult.
FARO-Droid is a new system that uses AI to detect malware (malicious software) on Android phones by analyzing multiple types of code features and combining them intelligently, even when attackers try to hide the malware through obfuscation (code transformation techniques that make programs harder to read and analyze). The system is designed to be reliable and resistant to these hiding techniques.
This academic article examines how hospital staff adopt defensive cybersecurity practices, focusing on the influence of leadership, human-centered skills, and rule-following requirements. The research explores organizational and behavioral factors that help healthcare workers protect systems and data from security threats, rather than technical fixes alone.
This academic paper describes FastPoS, a new security method for verifying that data is actually stored in fog-cloud IoT systems (networks of internet-connected devices distributed between local edge servers and central cloud storage). The method uses polynomial commitments (a cryptographic technique that lets someone prove they're storing data without revealing the data itself) to make verification faster and more efficient than existing approaches.
This research addresses leaky private information retrieval (L-PIR), a system where a user queries a database while accepting some measurable privacy leakage, quantified using differential privacy (a mathematical framework that bounds how much information an observer can learn about individual data). The researchers developed an improved L-PIR scheme that reduces privacy leakage from O(K) to O(log K) by using cyclic permutations (mathematical rearrangements) and assigning higher probabilities to keys with lower Hamming weights (fewer 1-bits in binary representation), achieving better privacy-to-cost tradeoffs than previous methods.
Model extraction attacks (MEA, where attackers steal the functionality of AI models by creating a clone with similar behavior) are a security threat that defenders counter by using auxiliary data to make the victim model give misleading predictions. However, realistic auxiliary data is hard to obtain, gives inconsistent protection, and doesn't protect all data categories equally. The paper proposes MDV (Model Defense Variational Autoencoder, a machine learning technique that generates synthetic data rather than using real data) to create virtual auxiliary data that effectively addresses all three problems.
Fix: The proposed solution is to use Model Defense Variational Autoencoder (MDV) to generate virtual auxiliary data as a replacement for realistic auxiliary data. MDV combines a Variational Autoencoder (VAE, a machine learning model that generates new synthetic data similar to training data) and a classifier, forcing learned features to follow different statistical distributions by category, then samples synthetic data from low-likelihood regions of these distributions to use as auxiliary data in defense methods.
IEEE Xplore (Security & AI Journals)Vul-CTG is a new AI framework for detecting software vulnerabilities (bugs that create security weaknesses) by combining two approaches: PLMs (pretrained language models, AI systems trained on large amounts of text) and GNNs (graph neural networks, AI systems that analyze connected data structures). The framework improves on existing methods by better combining code text analysis with program graph analysis, using contrastive learning (training the AI to recognize similarities and differences) and handling unreliable training labels, achieving about 3% better accuracy than previous approaches.
Researchers have developed a new backdoor attack method called Trigger without Trace (TwT) that can secretly compromise text-to-image diffusion models (AI systems that generate images from text descriptions) while avoiding detection. The method works by using syntactic structures (grammar patterns) as hidden triggers and employing a mathematical technique called Kernel Maximum Mean Discrepancy (KMMD, a way to match statistical distributions) to make malicious samples look identical to legitimate ones, achieving a 97.5% success rate while bypassing three existing defense detection systems.
This research presents a method to detect deepfakes (AI-generated fake videos or images of faces) by identifying inconsistencies in how image quality degrades between the background and the manipulated face regions. The approach uses a framework that learns to spot these degradation differences through two connected neural networks (deep learning models), one that creates fake images and another that detects them, working together in an adversarial process similar to a GAN (generative adversarial network, where two AI systems compete to improve each other). The method shows better performance when detecting deepfakes created by new, unseen manipulation techniques.