Academic papers, new techniques, benchmarks, and theoretical findings in AI/LLM security.
RAG (retrieval-augmented generation, where an AI pulls in external documents to answer questions) systems that connect LLMs to sensitive databases risk leaking private information during both the retrieval phase, where embeddings (numerical representations of text) can be reversed to recover original content, and the generation phase, where raw passages are exposed to the model. The paper proposes DP2-RAG, a framework using differential privacy (a mathematical technique that adds strategic noise to protect individual data while keeping results useful) to protect both stages: it uses noise-aware retrieval with correction to maintain accuracy while adding privacy-protecting noise, and a dual utility-exponent mechanism to protect generated text while keeping semantic meaning intact.
Fix: DP2-RAG introduces two mechanisms: (1) Noise-Aware Retrieval with Correction (NARC) enforces chunk-level differential privacy by adding calibrated noise and correcting ranking bias to mitigate accuracy degradation, and (2) Dual Utility-Exponent Mechanism (DUEM) guarantees token-level differential privacy for generated surrogates through a two-layer exponential mechanism while maintaining semantic fidelity.
IEEE Xplore (Security & AI Journals)This research presents a framework for sharing data while protecting privacy by carefully choosing which information to share based on specific privacy and utility goals. Instead of sharing all data or fixed subsets indiscriminately, the approach uses optimization (finding the best solution within constraints) to identify only the most useful data that meets privacy requirements, reducing unnecessary information leakage while maintaining data usefulness for tasks.
This paper presents RLAgent-GSSTI, a framework that uses reinforcement learning (RL, a machine learning technique where a system learns by receiving rewards for good actions) to automatically detect SSTI vulnerabilities (server-side template injection, where attackers manipulate template engines to execute unintended code on web servers). The framework combines code analysis tools with AI agents to both predict SSTI risks and generate attack payloads to identify vulnerabilities, achieving much lower false negative rates (missed vulnerabilities) compared to traditional security scanning tools.
This research introduces BG-FlipIn, a Bayesian game framework (a mathematical model where players make decisions based on incomplete information about each other) for studying advanced persistent threats (APT, long-term cyberattacks where an attacker maintains hidden access to a system) that involve an insider threat. The framework models three types of insiders—malicious, inadvertent, and corrupt—and helps defenders make strategic decisions about how often to check for or remove threats without needing to know exactly what the insider wants or constantly changing their defense approach.
This research proposes STD-INN (Spatial-Temporal Decoupled Invertible Neural Networks), a new AI method for detecting anomalies in skeleton-based video, which works by separating body pose information from motion patterns rather than processing them together. The approach uses normalizing flows (a type of neural network architecture) to analyze these two components independently: spatial flows examine unusual body positions using skeletal structure, while temporal flows detect irregular movement patterns using sequential analysis. The method achieves better performance than existing approaches while using fewer parameters and providing clearer explanations of why anomalies are detected.
Trilobyte is a system that hides encrypted data inside saved game files from single-player games, then shares this data between users through cloud-synchronized gaming accounts as a way to communicate secretly in heavily censored countries. The system can hide up to 5.3 MB of data per hour of gameplay and allows users to remove traces of communication after each use, making it harder to detect even if someone physically inspects a device or controls the gaming platform.
Researchers created R3S, a framework that improves how security experts use knowledge graphs (databases that show connections between related concepts) to predict defenses against cyber attacks. The framework combines two techniques: H-AIMG, which organizes information about vulnerabilities and attack methods, and DPSR, which merges two types of analysis (semantic, focused on meaning, and structural, focused on relationships) to better recommend mitigation measures for security threats.
Quantum-KIP is a method that compresses training data (the examples a machine learning model learns from) into a smaller set of representative points with adjusted labels, using quantum feature maps (functions that encode data using quantum computing). The method avoids backpropagation through quantum circuits (a computationally expensive process), and includes analysis showing that the compression provides privacy benefits by limiting how much changing one training example affects the model's predictions, while remaining robust to quantum noise (errors from imperfect quantum measurements).
ASFuzz is a testing tool that finds concurrency bugs (problems caused by multiple tasks running at the same time) in the Linux kernel, specifically bugs involving asynchronous tasks (operations that happen in the background without waiting for a response). The tool uses a Shadow Dependency Graph (a map of how different tasks relate to and trigger each other) and a dual-mode approach to explore both normal execution paths and stress-test specific interactions between synchronous (waiting) and asynchronous operations. ASFuzz discovered 14 previously unknown bugs in Linux kernel versions 6.6 to 6.15, including 10 asynchronous concurrency bugs, with 5 assigned CVE identifiers (official security vulnerability numbers).
This paper presents LOVA, a framework that improves how AI models find vulnerable code (code with security weaknesses) by using self-attention mechanisms (the components that help AI models figure out which parts of input text are most important). The key idea is that vulnerable lines of code will receive higher attention weights from the model, allowing LOVA to pinpoint security issues more accurately across different programming languages and achieve significantly better performance than existing AI-based approaches.
Researchers discovered that attackers can fool distributed AI systems (where neural networks are split across edge devices and cloud servers) by only having access to the edge portion. They created universal adversarial perturbations (tiny, crafted changes to input data designed to fool AI models), which can manipulate the feature representations (the internal data the model creates to understand images) at the edge device in ways that cause incorrect predictions even in the unseen cloud portion of the model. This attack works without the attacker knowing anything about the cloud component, showing a new security weakness in split AI systems.
Researchers developed CCCC, a method for hiding secret messages within standard TLS 1.3 connections (the protocol that encrypts web traffic) in a way that appears completely normal to network auditors and censors. The technique uses collision ciphertexts (specially crafted encrypted data that can be decrypted to different messages depending on which secret key is used) and embeds hidden information inside digital signatures, so an eavesdropper with only the public session key sees innocent messages while someone with a pre-shared secret key can extract the hidden communication.
Researchers discovered that FRAST, a cipher designed to work efficiently with TFHE (Torus-based Fully Homomorphic Encryption, a method for computing on encrypted data), is vulnerable to differential fault attacks (DFAs, where attackers deliberately introduce errors into a system to extract secret keys). The attack can recover the encryption key in seconds using just a few faults, making it the first successful fault-based attack against this cipher.
Fix: The source recommends two countermeasures: 'removing the negacyclic restriction in the penultimate round of FRAST and introducing non-zero linear structures into the S-boxes (substitution boxes, which scramble data) of the last two rounds.' The source notes that standard linear structures cannot be added to negacyclic S-boxes without breaking their efficiency in TFHE.
IEEE Xplore (Security & AI Journals)This paper presents MGEGOR, a new AI method for specific emitter identification (SEI, the process of authenticating wireless devices by analyzing their unique transmission characteristics). The method improves on existing approaches by better identifying both known devices seen during training and unknown devices that were not part of the training data, which is important for security in open-set conditions (scenarios where new, unauthorized devices may appear). The framework uses contrastive representation learning (a technique where the AI learns by comparing similar and dissimilar examples) and prototype-based embedding (storing representative examples of device types) to work effectively even when conditions change over time.
This paper proposes a routing system for uncrewed aerial vehicles (UAVs, or drones) that combines reinforcement learning (RL, a type of AI that learns by trial and error) with large language models (LLMs) to help drones find reliable communication paths when facing jamming attacks (deliberate radio interference) and gray-hole attacks (where nodes pretend to forward data but secretly drop it). The system uses sensor data like images and temperature readings to predict jamming and avoid compromised routes, resulting in better packet delivery and lower delays.
This research paper presents a comprehensive analysis of privacy threats in smart home systems, which are networked devices (like cameras, thermostats, and speakers) that collect data about your home and habits. The authors develop a privacy threat analysis (PTA) framework to identify and categorize different ways your personal information could be exposed or misused in these systems, and propose a risk management approach to help reduce those dangers. The work is intended to help engineers and developers build smarter homes that better protect residents' privacy.
This paper presents REI-Net, a new AI network that quickly extracts internal fingerprints (fingerprint patterns beneath skin) and internal sweat pores from OCT (optical coherence tomography, a medical imaging technique that captures 3D images below the surface). The method is much faster than previous approaches, completing analysis in 605 milliseconds, and achieves better accuracy for biometric security applications.
Ransomware (malicious software that encrypts files to extort money) has become more sophisticated by spreading attacks across multiple programs, making traditional process-level detection ineffective. This research presents CoLPR, a machine learning framework that learns to detect ransomware by understanding which behaviors are truly malicious versus which are just normal variations from different applications running together, achieving 100% detection success rate with only a 5-second delay.
This research paper presents a gait authentication framework, a security system that identifies people based on how they walk. The framework focuses on swing phase dynamics (the movement patterns during the part of a walking cycle when one leg swings forward), using AI to analyze these patterns for authentication purposes. The paper appears to be a technical study on using biometric walking patterns as a security method, published in a digital threats research journal.
This academic paper discusses a method for collecting data from multiple dimensions (different types of information) while protecting privacy using interval-based perturbation (adding controlled randomness to specific ranges of values) under differential privacy (a mathematical framework that limits how much an AI system can learn about individual data points). The research focuses on how to gather useful information while maintaining privacy guarantees.