Academic papers, new techniques, benchmarks, and theoretical findings in AI/LLM security.
This is a survey paper that reviews different frameworks, principles, and methods for auditing AI systems, published in an academic journal. The paper examines current approaches to evaluating and assessing AI systems for safety, fairness, and reliability, but does not focus on a specific vulnerability or security problem.
This research paper benchmarks (tests and compares) post-processing methods in local differential privacy (a technique that adds noise to data before it leaves a user's device to protect individual privacy) to evaluate how well they preserve data usefulness while resisting adversarial robustness (withstanding attacks designed to fool AI systems). The study measures different approaches to see which ones best balance keeping the data useful for analysis while protecting against attempts to reverse-engineer or attack the privacy mechanism.
This academic paper describes a method for hiding secret data inside video files using HEVC (High Efficiency Video Coding, a modern video compression standard) by modifying motion vectors (the data that describes how objects move between video frames). The technique uses a local optimality maintenance approach to keep the hidden data undetectable while maintaining video quality.
This academic paper proposes verifiable timed-release adaptor signatures, a cryptographic technique designed to improve security in blockchain-based auction systems. The method combines adaptor signatures (a way to link signatures across different transactions) with time-release mechanisms to ensure bids remain hidden until a specific moment, preventing unfair advantages in online auctions.
Researchers have developed a watermarking method for screenshot protection that uses a symmetric encoder-decoder system (a pair of matching neural networks that encode and decode information) combined with style transfer-based distortion modeling (applying artistic transformations to hide the watermark). This approach aims to make watermarks more resistant to attempts to remove them from screenshots while maintaining image quality.
This paper presents a new image encryption method that uses a 3D hyperchaotic map (a mathematical system with unpredictable behavior in three dimensions) combined with non-uniform block-based hybrid diffusion (mixing image data in varying-sized chunks to scramble pixel values). The research proposes a cryptographic technique designed to protect images from unauthorized access by making them appear as meaningless noise.
This research paper proposes a method for checking whether stored data remains unchanged and uncorrupted, while also removing duplicate copies to save storage space and protecting who owns which data. The approach combines data integrity auditing (verification that files haven't been modified), deduplication (eliminating redundant copies), and privacy protection so that ownership information stays confidential.
This academic paper examines whether analyzing the memory (the temporary data storage a program uses while running) of Android applications can effectively detect malicious behavior. The research, published in December 2026, investigates techniques for identifying harmful apps by studying what data and operations are stored in memory during execution.
This research paper presents BXLL, a system that combines Bi-LSTM (a type of neural network that processes data in both directions to recognize patterns) stacking with SDN-driven honeypots (fake systems designed to attract and study attackers) to detect unauthorized access attempts in IoT (internet of things, networks of connected devices) networks. The approach uses an ensemble method (combining multiple AI models to make better predictions) to improve the accuracy of identifying when someone is trying to break into IoT systems.
This academic paper proposes using blockchain (a distributed ledger technology where transactions are recorded in linked blocks that are difficult to tamper with) to create verifiable certificates and audit trails when judicial systems are unavailable. The research explores how blockchain's tamper-evident properties could help maintain trust and accountability in legal documentation even when court systems are down or inaccessible.
This academic paper describes a security system for sharing medical data among multiple people using a TEE (trusted execution environment, a secure area of a computer processor that protects sensitive operations). The system allows authorized users to have access removed (attribute revocation) to shared medical data while keeping the data protected from unauthorized access.
Researchers developed SSH-Tracker, a tool that analyzes SSH (Secure Shell, the encrypted protocol used to securely access remote computers) behavior by examining network traffic patterns without decrypting the encrypted communication itself. The tool can identify specific SSH activities and behaviors by studying how data flows through encrypted connections, potentially revealing user actions even when the actual content is protected.
This research paper presents MSTE-CAN, a deep learning model designed to detect unauthorized access attempts on CAN bus systems (the communication network used in vehicles and industrial equipment). The model uses multi-channel spatial-temporal encoding (a technique that analyzes patterns across multiple data streams over time) and coordinate attention mechanisms (methods that help the AI focus on important parts of the data) combined with ResNet (a type of neural network architecture) to identify intrusions more efficiently than previous approaches.
This academic paper analyzes security vulnerabilities in UEFI (Unified Extensible Firmware Interface, the low-level software that runs before your operating system starts). The research examines how attackers could exploit weaknesses at this firmware level, which is below the OS and therefore harder to detect and protect. Since UEFI runs before any security software can load, compromises at this layer pose serious risks to system security.
This research paper examines how malicious insiders (employees with legitimate system access who act with harmful intent) can exfiltrate data (steal or copy information), evade detection, and execute sequences of harmful actions, based on observations from a controlled behavioral experiment. The study was conducted by researchers at MITRE Corporation and provides evidence about the methods and patterns insiders use to avoid discovery while carrying out attacks.
This academic paper describes INFER, a portable hands-on lab system designed to teach digital forensics (the process of recovering and analyzing data from computers and devices to investigate crimes or incidents) to university students. The paper covers how INFER was designed, deployed in educational settings, and what impact it had on student learning outcomes.
This is a review paper examining cybersecurity exercises and evaluations that teach adversarial thinking (the skill of imagining how attackers might exploit systems). The paper analyzes different educational approaches and methods used to help students develop security mindsets by thinking like potential attackers.
This academic paper discusses training methods for responding to GNSS spoofing (a type of cyberattack where fake GPS signals trick navigation systems into showing incorrect locations) using ship-shore drills, which are coordinated practice exercises between vessels and coastal facilities. The paper, published in Computers & Security, examines how organizations can prepare personnel to detect and respond to GPS-based attacks through structured training scenarios.
This research paper discusses how Rust (a programming language designed to prevent memory errors) can be isolated from unsafe languages on ARM processors (a type of computer chip commonly used in phones and embedded devices) using hardware-based protection mechanisms. The work addresses the challenge of safely combining Rust with older or less safe programming languages in systems that need both, leveraging ARM's built-in security features to keep them separate.
This academic paper presents AdaCap-IIoT, a reference architecture that combines adaptive blockchain (a distributed ledger that can adjust its rules and processes) with CapBAC (capability-based access control, a security method where users get specific permissions for particular resources) designed for Industrial IoT (internet-connected devices in manufacturing and industrial settings). The work aims to provide a framework for securing industrial systems by combining adaptive blockchain technology with flexible permission management.