Academic papers, new techniques, benchmarks, and theoretical findings in AI/LLM security.
This research paper from November 2026 examines threat models for industrial control systems (ICS, the computers that manage factories, power plants, and other critical infrastructure) by developing systematic methods to search for and identify security threats. The authors appear to connect threat identification approaches with pentesting (penetration testing, where security experts deliberately try to break into systems to find weaknesses). The paper contributes to understanding how to better protect critical infrastructure from cyberattacks.
This academic paper presents ODHD, a method for generating helper data that enables reliable key derivation from SRAM PUF (static random-access memory physical unclonable function, a hardware feature that extracts unique cryptographic keys from the inherent variations in memory chips) without needing non-volatile memory (permanent storage like flash drives). The approach addresses the challenge of creating stable, reproducible keys from noisy hardware sources for secure cryptographic applications in resource-constrained devices.
This academic review examines intrusion detection systems (IDS, software that monitors networks to catch unauthorized access attempts) designed specifically for Internet of Medical Things (IoT devices like connected medical equipment that collect and share health data). The paper analyzes these systems across four key areas: how well they catch attacks, how efficiently they run, whether humans can understand why they flag something as a threat, and whether they work reliably on new types of attacks they haven't seen before.
Researchers discovered a new type of backdoor attack (hidden malicious code inserted into AI systems) that works against personalized federated learning (a privacy-focused method where multiple computers train an AI model together without sharing raw data). The attack is designed to be stealthy and persistent, meaning it can hide from detection and remain in the system over time.
BAVote is a blockchain-based electronic voting system that uses cryptographic techniques to balance voter privacy (hiding voter identity) with accountability (ability to trace dishonest voters). The paper presents a new approach called ConsATS, a threshold signature scheme (a cryptographic method where multiple parties must cooperate to create a valid signature) that compresses all voters' public keys into a single key, reducing storage needs by over 90% and making verification 7 times faster than previous systems.
Fix: The paper proposes using a new threshold signature scheme named ConsATS that features a constant-size verification key. This scheme compresses all voters' public keys into a single verification key, allowing aggregated ballots to be verified without storing or processing individual voter public keys. BAVote additionally combines one-time addresses (temporary, single-use identifiers) and a commit-reveal mechanism (a two-phase protocol where voters first commit to their choice, then later reveal it) to protect intermediate voting data on the blockchain, with a tracing key enabling authorized auditors to identify malicious voters.
IEEE Xplore (Security & AI Journals)NCorr-FP is a data fingerprinting (embedding hidden recipient-specific marks into data to track ownership) technique designed to protect structured tabular data in shared environments. The method preserves the statistical quality of data while embedding fingerprints using local record similarity and density estimation, and can detect these fingerprints even after substantial data modification or deletion attacks.
This research paper presents a framework for understanding how reinforcement learning-based cyber agents (AI systems trained to make decisions by trial and error in cybersecurity contexts) make their decisions. The authors developed a multi-layer approach to explain the "black box" problem (the difficulty in understanding why AI systems reach certain conclusions), which is important for security experts to verify that these AI agents are operating correctly and safely.
Three cybersecurity experts discuss why phishing training programs, which teach employees to recognize fake emails designed to steal information, often fail to change how people behave. They emphasize the need for better ways to measure whether training actually works and suggest strategies to help organizations defend against phishing attacks both now and in the future.
Debapt is a system that uses multiple AI agents (independent AI programs that work together) debating with each other to build profiles of APT (advanced persistent threat, a sophisticated type of cyberattack) adversaries by analyzing cyber threat intelligence (information about security threats). The system uses an ontology (a formal structure that defines how concepts relate to each other) to organize this debate process. This research proposes a new way to understand and track advanced attackers by having AI agents discuss and reason through threat data together.
This research paper presents a method for detecting deepfakes (synthetic videos or images created by AI to look realistic) using a dual-mode Swin Transformer, which is a type of neural network architecture. The approach uses multi-scale feature learning (analyzing visual details at different zoom levels) and local ambiguity mitigation (reducing confusion in uncertain areas) to improve detection accuracy. This is a technical contribution to security research, not a response to an existing vulnerability or security incident.
This research addresses a privacy risk in teleoperated robotics (systems where humans remotely control robots by having their movements tracked and converted into robot commands). The problem is that motion-tracking data can leak biometric information (unique physical characteristics) that could allow someone to re-identify the operator. The authors propose using a VAE (variational autoencoder, a type of machine learning model that learns compressed representations of data) to filter out identity-revealing patterns while keeping the motion information needed for the robot to complete tasks.
QS-BTrust is a new security protocol designed for Integrated Vehicular Networks (IVNs, which are connected vehicle communication systems) that authenticates broadcast messages while resisting attacks from quantum computers. The protocol combines Physical Unclonable Functions (PUFs, unique digital fingerprints built into hardware), post-quantum digital signatures (cryptographic techniques that remain secure even with quantum computers), and a Hashgraph-based system to verify messages with low overhead and support revoking compromised vehicles without slowing down traffic.
Researchers discovered a new security vulnerability in graph-based RAG (retrieval-augmented generation, where an AI system pulls information from external knowledge graphs to answer questions) systems used with large language models. Attackers can poison the external database by inserting hidden triggers and false information into the knowledge graph, causing the AI to give wrong answers when those triggers appear in user queries while still answering normal questions correctly. The attack uses three types of triggers at different complexity levels, from simple words to semantic patterns, and tests showed the attack works across multiple AI systems.
Physical adversarial attacks (PAAs, carefully crafted materials or objects that trick radar systems into giving wrong readings) threaten mmWave radar (millimeter-wave radar, a type of sensor used in autonomous vehicles and security systems) by manipulating its signals, but detecting these attacks has been difficult. mmGuard is a defense framework that identifies adversarial attacks by looking for three telltale physical signatures: spatial phase discontinuities (unnatural patterns in how radar waves reflect), anomalous radar cross-section patterns (unusual reflections), and violations of natural physics-based relationships, achieving over 90% detection accuracy.
Fix: mmGuard addresses the threat through multi-domain feature extraction to capture adversarial signatures, neural refinement to improve detection, and per-object attack detection and mitigation compatible with automotive radar update rates. The paper notes that few-shot adaptation enables calibration to unseen settings.
IEEE Xplore (Security & AI Journals)This research introduces QAFD (Quality-Assisted Forgery Detection), a new system for detecting AI-generated images by analyzing both visual features and quality-related artifacts that different generative models produce. The system uses a quality-guided approach to help AI models better understand degradation patterns in fake images, allowing it to detect AI-generated content more reliably even when tested on unseen generative models and images that have been edited after creation.
Ridesharing apps need to protect user location privacy, but adding random noise to locations (Laplace noise, a mathematical technique that obscures exact positions) makes it harder to match drivers with passengers efficiently. This paper proposes using linear programming (a mathematical optimization method for finding the best solution among many options) to solve the real-time matching problem between ridesharing requests and drivers while maintaining both privacy and matching quality.
Large language models raise privacy concerns because the knowledge they learn becomes deeply entangled in their structure, making it hard to make them "forget" specific information. Researchers developed a privacy-preserving machine unlearning method (a technique to remove learned data from AI models) that eliminates high-risk information from model outputs and uses differentially-private randomization (adding statistical noise to hide sensitive data) to ensure unlearned information cannot be identified, without requiring model parameter adjustments.
Fix: The proposed solution eliminates the impact of targeted information by removing high-risk semantic meanings from the model's output and incorporates differentially-private randomization to make the unlearned information statistically indiscernible. The algorithm requires neither parametric fine-tuning nor in-context prompt calibration.
IEEE Xplore (Security & AI Journals)This research proposes new cryptographic methods (PKEET-VPG and its verifiable version) to improve traceable anonymous credentials, which are systems that let users prove they have certain attributes without revealing their identity while still allowing regulators to trace misuse if needed. The key innovation uses session-specific tracing keys (temporary permission codes tied to single authentication sessions) to prevent abuse, and the methods reduce the computational burden on regulators who need to handle large numbers of authentication records.
This paper presents a mathematical framework for making cybersecurity decisions when facing uncertain threats in turn-based security games (scenarios where defenders and attackers take turns making moves). The framework handles uncertainty about both what the attacker might do and how well defensive controls will work, using game theory (the study of strategic decision-making between competing parties) and robust optimization (techniques for finding solutions that work well even when future conditions are unknown). The researchers show their approach outperforms traditional defensive strategies and demonstrate its usefulness through a network attack example.
This paper addresses the problem of detecting partially spoofed utterances (audio that contains both real and fake segments mixed together) without needing labeled data marking where the fake parts are. The researchers propose FMG, a method using Graph Neural Networks (GNNs, a type of AI model that understands relationships between connected pieces of data) to better track how different audio segments relate to each other over time and to identify which segments are likely fake.